[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8},"actionData","errors","root",{"_4509":4510,"_4511":-5,"_9":4512,"_4513":4514},"routes/library.video.$slug",{"_9":10,"_11":12,"_13":-5,"_14":15},"data",{"_16":17},"preview",false,"pageNumber","slug","goals-orchestration-and-the-human-control-plane","routeData",{"_18":19,"_20":21,"_22":23,"_24":25,"_26":27,"_28":29,"_30":31,"_32":-5,"_33":34,"_35":-5,"_36":37,"_38":39,"_40":41,"_42":43,"_44":-5,"_45":46,"_47":48,"_49":50,"_51":52,"_53":-5,"_54":-5,"_14":55,"_56":-5,"_57":58,"_59":60,"_61":62,"_63":64,"_65":66,"_25":67,"_68":69},"_createdAt","2026-10-23T21:54:05Z","_id","cda0904e-21aa-4a1d-b955-3f2f77be3480","_rev","ge7qSe37EQ3jcYkO7uARxV","_type","video","_updatedAt","2026-10-26T18:00:54Z","access","Public","afterEvent",true,"agenda","collections",[],"cta","endDate","2026-09-25T02:00:00.000Z","excerpt",[4502],"filmedOn","2026-09-24","library",{"_1993":1994,"_1995":1996},"modules","moreFromLibrary",[1984,1985,1986],"outlines",[1976],"posterFrame",{"_24":1973,"_74":1974},"publishDate","2026-10-26T15:15:00.000Z","skill","slideDeck",{"_24":14,"_1884":15},"speakers","speakersWithOrg",[1916],"startDate","2026-09-24T16:30:00.000Z","tags",[1869,1870,1871,1872,1873,1874],"title","Goals! Orchestration and the Human Control Plane","transcript",[151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185,186,187,188,189,190,191,192,193,194,195,196,197,198,199,200,201,202,203,204,205,206,207,208,209,210,211,212,213,214,215,216,217,218,219],{"_24":73,"_74":75},"seo",{"_70":71,"_72":50},"OGdescription","Paperclip cofounder Dotta on running a self-improving organization of agents: why skills without evals are like code without unit tests, why trust is an infrastructure problem, and why humans still set the goals, since agents can do anything you want but can't know what you value.","OGimage","mux.video","asset",{"_18":76,"_20":77,"_22":78,"_24":79,"_26":80,"_81":82,"_9":83,"_84":85,"_86":87,"_88":89,"_90":91},"2026-10-23T22:03:03Z","984c4e42-fa44-4a40-a0d2-f4ba9afa3ee0","YEARsfayN2fPM1bkAoMGQI","mux.videoAsset","2026-10-23T22:07:00Z","assetId","OmtnUN5fDhDMXJH7yVBgVLhRhs4xm5jA",{"_92":93,"_94":95,"_96":97,"_98":99,"_100":82,"_101":102,"_103":104,"_105":106,"_107":108,"_109":110,"_111":104,"_112":113,"_114":77,"_115":116,"_117":118,"_119":106,"_86":87,"_120":121,"_122":89,"_123":124},"playbackId","VuNqohmoUFpSv01sMexXFp1S6hWqpGdCA","status","ready","uploadId","l9EmpMXZVtNfRlgC9r8uTMEssAX01kVgC","url","https://stream.mux.com/VuNqohmoUFpSv01sMexXFp1S6hWqpGdCA","aspect_ratio","16:9","created_at","1792793012","duration",844.454233,"encoding_tier","smart","id","ingest_type","on_demand_direct_upload","master_access","none","max_resolution_tier","1080p","max_stored_frame_rate",29.97,"max_stored_resolution","UHD","mp4_support","non_standard_input_reasons",{"_149":150},"passthrough","playback_ids",[146],"progress",{"_117":143,"_144":145},"resolution_tier","tracks",[125,126],"upload_id","video_quality","plus",{"_96":135,"_100":136,"_137":138,"_139":140,"_141":31,"_134":142},{"_96":127,"_100":128,"_129":108,"_130":131,"_132":133,"_134":25},844.410233,"orf4oUvCP95XptZTovBYErs02E8nl9BZPQNmQdPGr9JE","max_frame_rate","max_height",1152,"max_width",2048,"type",844.410229,"JQjXAb1n6gvDEEqgVxbslSo8xSzu9Dew","max_channel_layout","stereo","max_channels",2,"primary","audio",100,"state","completed",{"_100":85,"_147":148},"policy","public","video_resolution","3840x2160",{"_220":1862,"_24":222,"_223":1863,"_225":1864,"_227":576},{"_220":1842,"_24":222,"_223":1843,"_225":1844,"_227":228},{"_220":1822,"_24":222,"_223":1823,"_225":1824,"_227":228},{"_220":1782,"_24":222,"_223":1783,"_225":1784,"_227":228},{"_220":1752,"_24":222,"_223":1753,"_225":1754,"_227":228},{"_220":1732,"_24":222,"_223":1733,"_225":1734,"_227":228},{"_220":1702,"_24":222,"_223":1703,"_225":1704,"_227":228},{"_220":1682,"_24":222,"_223":1683,"_225":1684,"_227":228},{"_220":1662,"_24":222,"_223":1663,"_225":1664,"_227":228},{"_220":1642,"_24":222,"_223":1643,"_225":1644,"_227":228},{"_220":1622,"_24":222,"_223":1623,"_225":1624,"_227":228},{"_220":1615,"_24":222,"_223":1616,"_225":1617,"_227":576},{"_220":1595,"_24":222,"_223":1596,"_225":1597,"_227":228},{"_220":1575,"_24":222,"_223":1576,"_225":1577,"_227":228},{"_220":1545,"_24":222,"_223":1546,"_225":1547,"_227":228},{"_220":1525,"_24":222,"_223":1526,"_225":1527,"_227":228},{"_220":1495,"_24":222,"_223":1496,"_225":1497,"_227":228},{"_220":1465,"_24":222,"_223":1466,"_225":1467,"_227":228},{"_220":1425,"_24":222,"_223":1426,"_225":1427,"_227":228},{"_220":1405,"_24":222,"_223":1406,"_225":1407,"_227":228},{"_220":1375,"_24":222,"_223":1376,"_225":1377,"_227":228},{"_220":1355,"_24":222,"_223":1356,"_225":1357,"_227":228},{"_220":1348,"_24":222,"_223":1349,"_225":1350,"_227":576},{"_220":1318,"_24":222,"_223":1319,"_225":1320,"_227":228},{"_220":1298,"_24":222,"_223":1299,"_225":1300,"_227":228},{"_220":1288,"_24":222,"_223":1289,"_225":1290,"_227":244},{"_220":1268,"_24":222,"_223":1269,"_225":1270,"_227":228},{"_220":1228,"_24":222,"_223":1229,"_225":1230,"_227":228},{"_220":1208,"_24":222,"_223":1209,"_225":1210,"_227":228},{"_220":1178,"_24":222,"_223":1179,"_225":1180,"_227":228},{"_220":1158,"_24":222,"_223":1159,"_225":1160,"_227":228},{"_220":1128,"_24":222,"_223":1129,"_225":1130,"_227":228},{"_220":1098,"_24":222,"_223":1099,"_225":1100,"_227":228},{"_220":1068,"_24":222,"_223":1069,"_225":1070,"_227":228},{"_220":1048,"_24":222,"_223":1049,"_225":1050,"_227":228},{"_220":1018,"_24":222,"_223":1019,"_225":1020,"_227":228},{"_220":998,"_24":222,"_223":999,"_225":1000,"_227":228},{"_220":991,"_24":222,"_223":992,"_225":993,"_227":576},{"_220":971,"_24":222,"_223":972,"_225":973,"_227":228},{"_220":951,"_24":222,"_223":952,"_225":953,"_227":244},{"_220":921,"_24":222,"_223":922,"_225":923,"_227":228},{"_220":901,"_24":222,"_223":902,"_225":903,"_227":228},{"_220":871,"_24":222,"_223":872,"_225":873,"_227":228},{"_220":851,"_24":222,"_223":852,"_225":853,"_227":228},{"_220":821,"_24":222,"_223":822,"_225":823,"_227":228},{"_220":801,"_24":222,"_223":802,"_225":803,"_227":228},{"_220":771,"_24":222,"_223":772,"_225":773,"_227":228},{"_220":741,"_24":222,"_223":742,"_225":743,"_227":244},{"_220":711,"_24":222,"_223":712,"_225":713,"_227":228},{"_220":691,"_24":222,"_223":692,"_225":693,"_227":228},{"_220":661,"_24":222,"_223":662,"_225":663,"_227":228},{"_220":641,"_24":222,"_223":642,"_225":643,"_227":228},{"_220":611,"_24":222,"_223":612,"_225":613,"_227":228},{"_220":581,"_24":222,"_223":582,"_225":583,"_227":228},{"_220":573,"_24":222,"_223":574,"_225":575,"_227":576},{"_220":543,"_24":222,"_223":544,"_225":545,"_227":228},{"_220":513,"_24":222,"_223":514,"_225":515,"_227":228},{"_220":493,"_24":222,"_223":494,"_225":495,"_227":228},{"_220":473,"_24":222,"_223":474,"_225":475,"_227":228},{"_220":453,"_24":222,"_223":454,"_225":455,"_227":228},{"_220":423,"_24":222,"_223":424,"_225":425,"_227":228},{"_220":403,"_24":222,"_223":404,"_225":405,"_227":228},{"_220":373,"_24":222,"_223":374,"_225":375,"_227":228},{"_220":343,"_24":222,"_223":344,"_225":345,"_227":228},{"_220":313,"_24":222,"_223":314,"_225":315,"_227":228},{"_220":293,"_24":222,"_223":294,"_225":295,"_227":228},{"_220":252,"_24":222,"_223":253,"_225":254,"_227":228},{"_220":241,"_24":222,"_223":242,"_225":243,"_227":244},{"_220":221,"_24":222,"_223":224,"_225":226,"_227":228},"_key","84b95bbfa9c3","block","children",[234],"markDefs",[229],"style","normal",{"_220":230,"_24":231,"_232":233},"7268245","timestamp","time","00:13:54.260",{"_220":235,"_24":236,"_237":238,"_239":240},"b0643b75f3b6","span","marks",[230],"text","So we're hiring at Paperclip if these issues interest you, and thank you very much.","7581598f150c",[248],[245],"blockquote",{"_220":246,"_24":231,"_232":247},"69d2327","00:13:48.019",{"_220":249,"_24":236,"_237":250,"_239":251},"b22c7841345e",[246],"And I would just leave you with this final thought, that agents can do anything that you want, but they can't know what you value.","4b10cbd2b81a",[267,268,269,270,271,272,273],[255,256,257,258],{"_220":265,"_24":231,"_232":266},{"_220":263,"_24":231,"_232":264},{"_220":261,"_24":231,"_232":262},{"_220":259,"_24":231,"_232":260},"2dfdcbc","00:13:44.779","abd184b","00:13:38.540","d4c74f2","00:13:34.360","1331182","00:13:31.399",{"_220":290,"_24":236,"_237":291,"_239":292},{"_220":288,"_24":236,"_237":289,"_239":279},{"_220":285,"_24":236,"_237":286,"_239":287},{"_220":283,"_24":236,"_237":284,"_239":279},{"_220":280,"_24":236,"_237":281,"_239":282},{"_220":277,"_24":236,"_237":278,"_239":279},{"_220":274,"_24":236,"_237":275,"_239":276},"62314a8409cf",[259],"This is something that we are building in Paperclip.","9d55f9566f7e",[]," ","4c4819c44a45",[261],"and essentially an operating system under the hood for the substrate for these agents to run in.","5323097c6fcc",[],"1e88d82f7763",[263],"You want to have an org chart for your agents,","856933f0982a",[],"42fb6156d16b",[265],"You want to have agent employee training.","ca77eed379c0",[302,303,304],[296,297],{"_220":300,"_24":231,"_232":301},{"_220":298,"_24":231,"_232":299},"451aa09","00:13:26.759","84f3f18","00:13:21.660",{"_220":310,"_24":236,"_237":311,"_239":312},{"_220":308,"_24":236,"_237":309,"_239":279},{"_220":305,"_24":236,"_237":306,"_239":307},"5529f6bf1d97",[298],"So really, four pieces that we see is you want a task manager to manage your agents.","02ef5addded1",[],"38aff9882c0b",[300],"And then you want to essentially have reflection over this whole process to make sure that you're able to improve it over time.","4f1134c1c143",[325,326,327,328,329],[316,317,318],{"_220":323,"_24":231,"_232":324},{"_220":321,"_24":231,"_232":322},{"_220":319,"_24":231,"_232":320},"a18dddc","00:13:12.440","00874aa","00:13:06.920","6cd7602","00:13:02.960",{"_220":340,"_24":236,"_237":341,"_239":342},{"_220":338,"_24":236,"_237":339,"_239":279},{"_220":335,"_24":236,"_237":336,"_239":337},{"_220":333,"_24":236,"_237":334,"_239":279},{"_220":330,"_24":236,"_237":331,"_239":332},"8b0e43a413b6",[319],"You want to have agents that criticize and give feedback and keep an eye on other agents to make sure that they're progressing properly and that they're doing the work properly.","0d035bce2569",[],"4f548eaad118",[321],"You want to have delegating agents which can then delegate work to other agents and receive results back.","0366c55c6305",[],"c20b9cf5ceb5",[323],"You want to be able to have teams of agents that you can assign work and get work back from them.","04ddde90348e",[355,356,357,358,359],[346,347,348],{"_220":353,"_24":231,"_232":354},{"_220":351,"_24":231,"_232":352},{"_220":349,"_24":231,"_232":350},"1e6c97f","00:12:58.160","8d9295a","00:12:46.600","850eed7","00:12:39.480",{"_220":370,"_24":236,"_237":371,"_239":372},{"_220":368,"_24":236,"_237":369,"_239":279},{"_220":365,"_24":236,"_237":366,"_239":367},{"_220":363,"_24":236,"_237":364,"_239":279},{"_220":360,"_24":236,"_237":361,"_239":362},"74143ca22844",[349],"basis. You want to be able to have roles for an agent that have these defined boundaries.","3ea60976b961",[],"142aa39d69c4",[351],"And I would say most teams are standing here, right? You assign tasks to your agent, you have skills to make sure it does the right thing, but on top of this you want to be able to schedule tasks that are occurring on a recurring","5fdc03ec15ca",[],"db8f6b508496",[353],"So when we look at what are these eight layers that you need for an organization that can learn, you have essentially tasks and skills.","78b995d018de",[385,386,387,388,389],[376,377,378],{"_220":383,"_24":231,"_232":384},{"_220":381,"_24":231,"_232":382},{"_220":379,"_24":231,"_232":380},"7b0c80a","00:12:34.080","66daa0e","00:12:26.800","220f716","00:12:20.240",{"_220":400,"_24":236,"_237":401,"_239":402},{"_220":398,"_24":236,"_237":399,"_239":279},{"_220":395,"_24":236,"_237":396,"_239":397},{"_220":393,"_24":236,"_237":394,"_239":279},{"_220":390,"_24":236,"_237":391,"_239":392},"9ead358f00e5",[379],"Routines would be documented processes. You have program managers, you have managing other managers.","55dc7df8eda1",[],"adfa96a97e7c",[381],"Really a lot of the management processes that we've already known from traditional organizations still apply to agents, right?","fde15b30d3ba",[],"dcfbb6ba8855",[383],"One, your job is to find the algorithm of your organization, and the work is in defining what's good enough and measuring it.","c6b98fd75858",[412,413,414],[406,407],{"_220":410,"_24":231,"_232":411},{"_220":408,"_24":231,"_232":409},"44d2829","00:12:17.840","466ad91","00:12:11.480",{"_220":420,"_24":236,"_237":421,"_239":422},{"_220":418,"_24":236,"_237":419,"_239":279},{"_220":415,"_24":236,"_237":416,"_239":417},"de56600ff3b4",[408],"So here are a few parting thoughts for you.","67ab8dc29415",[],"6245c2c65407",[410],"And so again, we're back to the same problem, which is the two hardest problems are defining what is good and what is actually done.","02dbda182d40",[435,436,437,438,439],[426,427,428],{"_220":433,"_24":231,"_232":434},{"_220":431,"_24":231,"_232":432},{"_220":429,"_24":231,"_232":430},"9119abb","00:12:06.680","4cddaff","00:11:59.260","ae704a5","00:11:55.019",{"_220":450,"_24":236,"_237":451,"_239":452},{"_220":448,"_24":236,"_237":449,"_239":279},{"_220":445,"_24":236,"_237":446,"_239":447},{"_220":443,"_24":236,"_237":444,"_239":279},{"_220":440,"_24":236,"_237":441,"_239":442},"30c18d567ed3",[429],"These things are easy, but measuring kind of the value on the other end is also very, very hard.","0d70c2493688",[],"537e8f65cf7d",[431],"Measuring the cost of the blog post and measuring the views that it got and measuring the conversions, right?","001bd3bcaf97",[],"17f61c0e410e",[433],"You can say, well, I measure the cost of this blog post and how many sales came out of it.","92db236cf9a1",[462,463,464],[456,457],{"_220":460,"_24":231,"_232":461},{"_220":458,"_24":231,"_232":459},"88dbb64","00:11:50.820","02ff01c","00:11:39.620",{"_220":470,"_24":236,"_237":471,"_239":472},{"_220":468,"_24":236,"_237":469,"_239":279},{"_220":465,"_24":236,"_237":466,"_239":467},"c6325e02d4bd",[458],"And that said, I think measuring the token cost, right?","6d2dd0fd471d",[],"add5fd1480ac",[460],"I think that one of the things that's really interesting with agents is we are now able to actually run that company equation with a lot more detail than we've ever been able to before.","459c7840d1c8",[482,483,484],[476,477],{"_220":480,"_24":231,"_232":481},{"_220":478,"_24":231,"_232":479},"93bd162","00:11:33.620","5c8ac6a","00:11:26.899",{"_220":490,"_24":236,"_237":491,"_239":492},{"_220":488,"_24":236,"_237":489,"_239":279},{"_220":485,"_24":236,"_237":486,"_239":487},"ec36e44a4668",[478],"And if you understand the equation of your company, then you are able to optimize it and essentially be successful.","0c71b983d9e8",[],"5ef592a7f8c4",[480],"So many of you might be familiar with this Keith Rabois quote, which he basically reduces your company down to an equation.","7f36f054cc06",[502,503,504],[496,497],{"_220":500,"_24":231,"_232":501},{"_220":498,"_24":231,"_232":499},"a99f1fd","00:11:24.420","d58788c","00:11:20.660",{"_220":510,"_24":236,"_237":511,"_239":512},{"_220":508,"_24":236,"_237":509,"_239":279},{"_220":505,"_24":236,"_237":506,"_239":507},"2add05444e2d",[498],"And you also need to specify a goal.","551225ccce30",[],"5ed54f388db0",[500],"So Paperclip gives you this, Buzz would be one, Centaur is another one.","c7ab204f2586",[525,526,527,528,529],[516,517,518],{"_220":523,"_24":231,"_232":524},{"_220":521,"_24":231,"_232":522},{"_220":519,"_24":231,"_232":520},"ee113e9","00:11:17.940","cae584c","00:11:14.420","b1dff7d","00:11:07.840",{"_220":540,"_24":236,"_237":541,"_239":542},{"_220":538,"_24":236,"_237":539,"_239":279},{"_220":535,"_24":236,"_237":536,"_239":537},{"_220":533,"_24":236,"_237":534,"_239":279},{"_220":530,"_24":236,"_237":531,"_239":532},"eba221d7b402",[519],"But you also need this kind of organizing structure.","d8b1e3da063a",[],"496bd2a5d994",[521],"rather than something maybe you're trying to do within your company.","0f06a97ed912",[],"656c79fe9bdf",[523],"And these tools here would be more for, let's say, autonomous organizations that are running completely on their own,","7402b9f11982",[555,556,557,558,559],[546,547,548],{"_220":553,"_24":231,"_232":554},{"_220":551,"_24":231,"_232":552},{"_220":549,"_24":231,"_232":550},"8401f25","00:11:05.440","c6b4cc9","00:11:03.940","a9d86ae","00:10:59.860",{"_220":570,"_24":236,"_237":571,"_239":572},{"_220":568,"_24":236,"_237":569,"_239":279},{"_220":565,"_24":236,"_237":566,"_239":567},{"_220":563,"_24":236,"_237":564,"_239":279},{"_220":560,"_24":236,"_237":561,"_239":562},"421ab03549dd",[549],"First off, you need this kind of substrate to run in.","b7def5ef14f4",[],"dcac2987c945",[551],"How do we actually do it?","0c898d8e2d20",[],"2a45cdf0f372",[553],"So if we're going to make this self-improving organization, how?","1a117a75a1e5",[577],[],"h2",{"_220":578,"_24":236,"_237":579,"_239":580},"04e8e15e4622",[],"Goals and the Eight Layers","7670756a4f13",[593,594,595,596,597],[584,585,586],{"_220":591,"_24":231,"_232":592},{"_220":589,"_24":231,"_232":590},{"_220":587,"_24":231,"_232":588},"56d3158","00:10:56.660","0c2fa90","00:10:51.680","bb3b1d6","00:10:43.819",{"_220":608,"_24":236,"_237":609,"_239":610},{"_220":606,"_24":236,"_237":607,"_239":279},{"_220":603,"_24":236,"_237":604,"_239":605},{"_220":601,"_24":236,"_237":602,"_239":279},{"_220":598,"_24":236,"_237":599,"_239":600},"a436abdd76b8",[587],"And you can't optimize without the proper evals, right?","f390d8518b7f",[],"9a9b916ffa1c",[589],"But essentially in order to learn that, you have to have an objective function that you're trying to optimize.","e1a654efee22",[],"f97f8e2146a1",[591],"I think eventually we probably will get to this sort of latent organizational pattern that the agents discover themselves.","4f6c3d56324c",[623,624,625,626,627],[614,615,616],{"_220":621,"_24":231,"_232":622},{"_220":619,"_24":231,"_232":620},{"_220":617,"_24":231,"_232":618},"6e79b63","00:10:41.019","0531b5c","00:10:37.819","296b90a","00:10:28.620",{"_220":638,"_24":236,"_237":639,"_239":640},{"_220":636,"_24":236,"_237":637,"_239":279},{"_220":633,"_24":236,"_237":634,"_239":635},{"_220":631,"_24":236,"_237":632,"_239":279},{"_220":628,"_24":236,"_237":629,"_239":630},"fd458d8d7cf0",[617],"You have a chief of staff that knows about the agents in your org.","a193dc156c2d",[],"b943b6ece5ac",[619],"one agent sort of routes to the other agents that it needs to, right?","770691077ab2",[],"8254f763d9fa",[621],"And I think that's a pattern that we're seeing, is not so much that it replicates exactly a human org chart, but rather that there is routing where","de233402841d",[650,651,652],[644,645],{"_220":648,"_24":231,"_232":649},{"_220":646,"_24":231,"_232":647},"c65d3ca","00:10:24.620","cccd065","00:10:19.020",{"_220":658,"_24":236,"_237":659,"_239":660},{"_220":656,"_24":236,"_237":657,"_239":279},{"_220":653,"_24":236,"_237":654,"_239":655},"3c62f89eeb71",[646],"So we certainly will see more of this kind of delegation agents.","6a1ce92a4893",[],"d7933edbd7a6",[648],"And there's probably really only so many agents that you can ever remember who they are and what they have.","8b0451e72a66",[673,674,675,676,677],[664,665,666],{"_220":671,"_24":231,"_232":672},{"_220":669,"_24":231,"_232":670},{"_220":667,"_24":231,"_232":668},"10f0df0","00:10:15.660","1ed43a8","00:10:11.800","1642625","00:10:09.000",{"_220":688,"_24":236,"_237":689,"_239":690},{"_220":686,"_24":236,"_237":687,"_239":279},{"_220":683,"_24":236,"_237":684,"_239":685},{"_220":681,"_24":236,"_237":682,"_239":279},{"_220":678,"_24":236,"_237":679,"_239":680},"9961e99eb2a5",[667],"what kind of permissions they have, what expectations that we have of their behavior.","cc74ba12b190",[],"92570bdab8d6",[669],"Even if it's the same model under the hood, we need to be able to know who we're talking to,","89a5b55d880d",[],"0e2ecedcee04",[671],"And ultimately, humans need a face.","f44e4cd45f02",[700,701,702],[694,695],{"_220":698,"_24":231,"_232":699},{"_220":696,"_24":231,"_232":697},"f5f5b08","00:10:00.280","13cd4ac","00:09:53.640",{"_220":708,"_24":236,"_237":709,"_239":710},{"_220":706,"_24":236,"_237":707,"_239":279},{"_220":703,"_24":236,"_237":704,"_239":705},"79f86037b44f",[696],"Speaking of incentives, also you want to have different agents that have different responsibilities.","7201f38b5d87",[],"dc05e309b805",[698],"There's more fine-tuned models, there's more open models, and they're gonna have different incentive systems.","a16ffe1664c5",[723,724,725,726,727],[714,715,716],{"_220":721,"_24":231,"_232":722},{"_220":719,"_24":231,"_232":720},{"_220":717,"_24":231,"_232":718},"057bef6","00:09:48.600","0133073","00:09:39.380","10c823c","00:09:36.740",{"_220":738,"_24":236,"_237":739,"_239":740},{"_220":736,"_24":236,"_237":737,"_239":279},{"_220":733,"_24":236,"_237":734,"_239":735},{"_220":731,"_24":236,"_237":732,"_239":279},{"_220":728,"_24":236,"_237":729,"_239":730},"234d285eec29",[717],"And so I think that we are only going to move into a world where that becomes more true.","084ffad67afe",[],"0cdbb4a525b7",[719],"You know that if you've used Grok or ChatGPT or Claude, that they speak differently, they have different kind of opinions, they'll tell you different things.","2271ac886127",[],"d576a9c22750",[721],"I also think that the different models have different personalities.","59ce834b1931",[753,754,755,756,757],[744,745,746],{"_220":751,"_24":231,"_232":752},{"_220":749,"_24":231,"_232":750},{"_220":747,"_24":231,"_232":748},"b846217","00:09:30.579","13351ed","00:09:22.399","ff007bd","00:09:18.080",{"_220":768,"_24":236,"_237":769,"_239":770},{"_220":766,"_24":236,"_237":767,"_239":279},{"_220":763,"_24":236,"_237":764,"_239":765},{"_220":761,"_24":236,"_237":762,"_239":279},{"_220":758,"_24":236,"_237":759,"_239":760},"b4d2daa68f08",[747],"has a lot of weight to the model where it understands how to operate, what the normative expectations are, and how it performs.","6decb021e1be",[],"31d2fce01e7b",[749],"When you tell a model, I want a security engineer who reports to the CTO, it already understands, you're invoking a magic spell that","d3c37c906739",[],"70045298c2f0",[751],"I think one thing is that the org charts are already built into the weights.","afa02c9a6866",[783,784,785,786,787],[774,775,776],{"_220":781,"_24":231,"_232":782},{"_220":779,"_24":231,"_232":780},{"_220":777,"_24":231,"_232":778},"3874b67","00:09:10.180","e205042","00:09:02.900","413eb8b","00:08:59.300",{"_220":798,"_24":236,"_237":799,"_239":800},{"_220":796,"_24":236,"_237":797,"_239":279},{"_220":793,"_24":236,"_237":794,"_239":795},{"_220":791,"_24":236,"_237":792,"_239":279},{"_220":788,"_24":236,"_237":789,"_239":790},"6fce02658082",[777],"And I'll make the case for you that actually, a lot of different agents is an important piece of making these self-learning organizations.","725dadad4651",[],"71b7539da1e4",[779],"Maybe we just need one God agent in our organization that does wish fulfillment, or do we actually have a lot of different agents in our org?","0823f24c9675",[],"8cee67895a65",[781],"There's this open question, which is like, do you need more than one agent?","73111e484c66",[810,811,812],[804,805],{"_220":808,"_24":231,"_232":809},{"_220":806,"_24":231,"_232":807},"e388770","00:08:56.720","a999606","00:08:49.760",{"_220":818,"_24":236,"_237":819,"_239":820},{"_220":816,"_24":236,"_237":817,"_239":279},{"_220":813,"_24":236,"_237":814,"_239":815},"f9b6223c41ec",[806],"Different agents need to have different access.","e3c56b39bbe5",[],"f7ef33fd463d",[808],"So this is another reason that it's important to have different agents within your organization.","ca92d837b3ad",[833,834,835,836,837],[824,825,826],{"_220":831,"_24":231,"_232":832},{"_220":829,"_24":231,"_232":830},{"_220":827,"_24":231,"_232":828},"1745dc6","00:08:47.420","09b4cba","00:08:43.099","95b7c87","00:08:35.280",{"_220":848,"_24":236,"_237":849,"_239":850},{"_220":846,"_24":236,"_237":847,"_239":279},{"_220":843,"_24":236,"_237":844,"_239":845},{"_220":841,"_24":236,"_237":842,"_239":279},{"_220":838,"_24":236,"_237":839,"_239":840},"b8608a12b6db",[827],"And this is totally open source.","2ab149d36c7b",[],"62f1ec6d5a2b",[829],"MCP skills and whether you want to deny it, allow it, or ask a human.","a03f1a26db87",[],"60bacab192fc",[831],"Arcade, Composio, Executor. Paperclip itself has this built in where you can control your individual","1cc030a6c83d",[860,861,862],[854,855],{"_220":858,"_24":231,"_232":859},{"_220":856,"_24":231,"_232":857},"5577850","00:08:32.880","5be6e4d","00:08:29.840",{"_220":868,"_24":236,"_237":869,"_239":870},{"_220":866,"_24":236,"_237":867,"_239":279},{"_220":863,"_24":236,"_237":864,"_239":865},"02c3932f52e9",[856],"And there are a number of permission-aware proxy tools.","491c4fed76e9",[],"84ce5503cc28",[858],"There's lots of technical ways that you can do it, but essentially this is a pattern that we're seeing.","10e7dd1e8ec1",[883,884,885,886,887],[874,875,876],{"_220":881,"_24":231,"_232":882},{"_220":879,"_24":231,"_232":880},{"_220":877,"_24":231,"_232":878},"cad2aaf","00:08:26.400","f25d281","00:08:17.720","ea04ddb","00:08:13.840",{"_220":898,"_24":236,"_237":899,"_239":900},{"_220":896,"_24":236,"_237":897,"_239":279},{"_220":893,"_24":236,"_237":894,"_239":895},{"_220":891,"_24":236,"_237":892,"_239":279},{"_220":888,"_24":236,"_237":889,"_239":890},"c625b72393ed",[877],"And you can do this either through an HTTP proxy or whatever.","58369b570731",[],"8a8d34384863",[879],"the permissions that they have access to are constrained by some other control plane that they are not, you know, short of any bugs, they're not able to sort of penetrate.","1aef811a2322",[],"d3ab95af842d",[881],"This idea is that","182d6a312104",[910,911,912],[904,905],{"_220":908,"_24":231,"_232":909},{"_220":906,"_24":231,"_232":907},"1235fda","00:08:04.140","80364ea","00:07:59.340",{"_220":918,"_24":236,"_237":919,"_239":920},{"_220":916,"_24":236,"_237":917,"_239":279},{"_220":913,"_24":236,"_237":914,"_239":915},"b5b724e63b16",[906],"You're losing kind of the tools, the keys, this persistent state, access to the world, and so one of the patterns that we see is having a permissions-aware proxy.","0da14e133c72",[],"6815ac965230",[908],"But the problem is as soon as you put your agent into a sandbox, you lose everything that made it powerful.","7453801c5b2a",[933,934,935,936,937],[924,925,926],{"_220":931,"_24":231,"_232":932},{"_220":929,"_24":231,"_232":930},{"_220":927,"_24":231,"_232":928},"bb2d5e7","00:07:53.900","f2ab08a","00:07:49.660","893c87f","00:07:44.580",{"_220":948,"_24":236,"_237":949,"_239":950},{"_220":946,"_24":236,"_237":947,"_239":279},{"_220":943,"_24":236,"_237":944,"_239":945},{"_220":941,"_24":236,"_237":942,"_239":279},{"_220":938,"_24":236,"_237":939,"_239":940},"f64196447f20",[927],"You know, XE, Unikraft, and a bunch of other providers provide sandboxes for your agents.","ee876e77f86f",[],"b541f791fd75",[929],"So one of the main ways that people do this with agents today would be sandboxing.","fc19fcd9fb76",[],"3438f0014938",[931],"If you don't trust giving an agent access within your org, it's because you don't have the infrastructure to contain it.","c86d64566668",[960,961,962],[954,955],{"_220":958,"_24":231,"_232":959},{"_220":956,"_24":231,"_232":957},"7a5e3e0","00:07:42.419","ef606e8","00:07:34.580",{"_220":968,"_24":236,"_237":969,"_239":970},{"_220":966,"_24":236,"_237":967,"_239":279},{"_220":963,"_24":236,"_237":964,"_239":965},"977681f3dc43",[956],"And so trust is an infrastructure problem.","7ccfb426f6d0",[],"6a5ff06a89f2",[958],"And one of the things you should know is that if you don't trust your agents, it means that you don't have a mechanism to verify it or constrain it.","77221a0ddbd9",[980,981,982],[974,975],{"_220":978,"_24":231,"_232":979},{"_220":976,"_24":231,"_232":977},"b815c50","00:07:31.700","c21ad4c","00:07:28.340",{"_220":988,"_24":236,"_237":989,"_239":990},{"_220":986,"_24":236,"_237":987,"_239":279},{"_220":983,"_24":236,"_237":984,"_239":985},"81660c7a9baa",[976],"In fact, it'll try to unlock the door even if you don't give it the key.","03383c9371a9",[],"181a8675d252",[978],"If you give your agent a key, it will unlock any door it possibly can.","27e0976af3f3",[994],[],{"_220":995,"_24":236,"_237":996,"_239":997},"29dd1a7f9ea4",[],"Trust and Agent Roles","84df4f9dab32",[1007,1008,1009],[1001,1002],{"_220":1005,"_24":231,"_232":1006},{"_220":1003,"_24":231,"_232":1004},"12734e1","00:07:20.940","894f92b","00:07:15.180",{"_220":1015,"_24":236,"_237":1016,"_239":1017},{"_220":1013,"_24":236,"_237":1014,"_239":279},{"_220":1010,"_24":236,"_237":1011,"_239":1012},"aac425862c56",[1003],"What can your agents write? One of the things we find with agents is it's not that they can't do enough right now, but they actually can do too much.","466fc48f04a6",[],"8bf96f615384",[1005],"We have that in Paperclip. So when you have these skills, you have this new problem, like, what can your agents read?","276678f915ad",[1030,1031,1032,1033,1034],[1021,1022,1023],{"_220":1028,"_24":231,"_232":1029},{"_220":1026,"_24":231,"_232":1027},{"_220":1024,"_24":231,"_232":1025},"e87f28d","00:07:11.540","8ad4927","00:07:08.340","76973e9","00:07:06.020",{"_220":1045,"_24":236,"_237":1046,"_239":1047},{"_220":1043,"_24":236,"_237":1044,"_239":279},{"_220":1040,"_24":236,"_237":1041,"_239":1042},{"_220":1038,"_24":236,"_237":1039,"_239":279},{"_220":1035,"_24":236,"_237":1036,"_239":1037},"1773bc174b67",[1024],"You should be using tools that help you develop and share skills.","02fd3ae2e611",[],"5fedfd17a184",[1026],"This is another tip. You should schedule agents to look for missing skills.","89a79c1aead2",[],"95a33e75cbc3",[1028],"And so you can discover skills from your traces.","dc023d1acee3",[1057,1058,1059],[1051,1052],{"_220":1055,"_24":231,"_232":1056},{"_220":1053,"_24":231,"_232":1054},"ee67c82","00:06:57.560","b381346","00:06:48.639",{"_220":1065,"_24":236,"_237":1066,"_239":1067},{"_220":1063,"_24":236,"_237":1064,"_239":279},{"_220":1060,"_24":236,"_237":1061,"_239":1062},"e8e44201d3a2",[1053],"Like, yes, you get a bit of improvement when new evals come out, but if you're aiming at the wrong thing, you're not going to get the results that you want.","1b48d9764937",[],"b09eb046d5f3",[1055],"And so this is like a third rung of more advanced evals. An important piece is that smarter models won't fix your bad directions.","36e78cb8dfcb",[1080,1081,1082,1083,1084],[1071,1072,1073],{"_220":1078,"_24":231,"_232":1079},{"_220":1076,"_24":231,"_232":1077},{"_220":1074,"_24":231,"_232":1075},"9d21e8b","00:06:45.039","05171db","00:06:33.000","8ab19d3","00:06:28.780",{"_220":1095,"_24":236,"_237":1096,"_239":1097},{"_220":1093,"_24":236,"_237":1094,"_239":279},{"_220":1090,"_24":236,"_237":1091,"_239":1092},{"_220":1088,"_24":236,"_237":1089,"_239":279},{"_220":1085,"_24":236,"_237":1086,"_239":1087},"3f7eb6fd4478",[1074],"are aiming for are actually better as a consequence.","5405fb898a78",[],"dff8b41127ce",[1076],"I have a coder and I have a security team and I have a QA team, and we want to understand, to make sure that the agents themselves are fulfilling their different roles and that the outcomes that they","57a1d67eb207",[],"27ed9851a8e8",[1078],"There's this emerging new set of evals, would be like organizational evals.","e3efc405afce",[1110,1111,1112,1113,1114],[1101,1102,1103],{"_220":1108,"_24":231,"_232":1109},{"_220":1106,"_24":231,"_232":1107},{"_220":1104,"_24":231,"_232":1105},"b1eeaa6","00:06:22.780","fcd0793","00:06:17.660","7826101","00:06:09.900",{"_220":1125,"_24":236,"_237":1126,"_239":1127},{"_220":1123,"_24":236,"_237":1124,"_239":279},{"_220":1120,"_24":236,"_237":1121,"_239":1122},{"_220":1118,"_24":236,"_237":1119,"_239":279},{"_220":1115,"_24":236,"_237":1116,"_239":1117},"5d8d56ebf2c2",[1104],"You need an agentic harness that actually runs through in order to get evals over this longer period of time.","8e24afe5f580",[],"f1f2c3b791b8",[1106],"I want you to write this piece of code, write an Express server, write the unit tests for it.","9d45793c5db6",[],"c1ce41328085",[1108],"Agentic evals actually need an agentic harness where you say like, follow these set of instructions and I need you to go across a trajectory, right?","ec0ca4bbb946",[1140,1141,1142,1143,1144],[1131,1132,1133],{"_220":1138,"_24":231,"_232":1139},{"_220":1136,"_24":231,"_232":1137},{"_220":1134,"_24":231,"_232":1135},"e0e680e","00:06:08.380","9c516ba","00:06:06.140","1c5b31e","00:06:01.340",{"_220":1155,"_24":236,"_237":1156,"_239":1157},{"_220":1153,"_24":236,"_237":1154,"_239":279},{"_220":1150,"_24":236,"_237":1151,"_239":1152},{"_220":1148,"_24":236,"_237":1149,"_239":279},{"_220":1145,"_24":236,"_237":1146,"_239":1147},"238b758fc815",[1134],"And you get a result there.","38267f57238b",[],"0dbfd387e664",[1136],"Like, what is the capital of France, right?","7f1c6a7cc9ec",[],"33d29959fff6",[1138],"There's model evals. A tool like promptfoo would say, I put a prompt in, I get a prompt out, right?","cda9ba6b1ec1",[1167,1168,1169],[1161,1162],{"_220":1165,"_24":231,"_232":1166},{"_220":1163,"_24":231,"_232":1164},"e1d8fdb","00:05:59.240","e6c8d1c","00:05:53.240",{"_220":1175,"_24":236,"_237":1176,"_239":1177},{"_220":1173,"_24":236,"_237":1174,"_239":279},{"_220":1170,"_24":236,"_237":1171,"_239":1172},"f09bc1f5e7b7",[1163],"There's really a ladder of evals too.","de3e99e807ae",[],"717e06fae71e",[1165],"So these are challenges, but it's really important that you understand investing in evals right now.","62a9e52f6d9f",[1190,1191,1192,1193,1194],[1181,1182,1183],{"_220":1188,"_24":231,"_232":1189},{"_220":1186,"_24":231,"_232":1187},{"_220":1184,"_24":231,"_232":1185},"4561a4a","00:05:48.060","f2eb6f3","00:05:45.180","0bd842d","00:05:42.040",{"_220":1205,"_24":236,"_237":1206,"_239":1207},{"_220":1203,"_24":236,"_237":1204,"_239":279},{"_220":1200,"_24":236,"_237":1201,"_239":1202},{"_220":1198,"_24":236,"_237":1199,"_239":279},{"_220":1195,"_24":236,"_237":1196,"_239":1197},"31503c13ad67",[1184],"And sometimes you'll spend a lot of time on a skill and the fundamental underlying models would just absorb them.","72aa98a8f502",[],"42e05b3df684",[1186],"They break across new models, they'll tend to over-optimize.","890be96606c5",[],"19cee0912337",[1188],"So of course the challenges with evals is that they take a long time.","e8e80cefc503",[1217,1218,1219],[1211,1212],{"_220":1215,"_24":231,"_232":1216},{"_220":1213,"_24":231,"_232":1214},"abdae1b","00:05:34.680","76942f4","00:05:23.220",{"_220":1225,"_24":236,"_237":1226,"_239":1227},{"_220":1223,"_24":236,"_237":1224,"_239":279},{"_220":1220,"_24":236,"_237":1221,"_239":1222},"f776585f48af",[1213],"So the savings here can be enormous because essentially if you know that your skills work well across a lot of dimensions, you can save a lot of cost.","ac1e4222cfd9",[],"ce09afbecf25",[1215],"Famously back in April, Shopify used DSPy and fine-tuned some of their own models to basically take an inference path that cost five million a year down to 73,000 a year.","2745e62cf548",[1243,1244,1245,1246,1247,1248,1249],[1231,1232,1233,1234],{"_220":1241,"_24":231,"_232":1242},{"_220":1239,"_24":231,"_232":1240},{"_220":1237,"_24":231,"_232":1238},{"_220":1235,"_24":231,"_232":1236},"bf69577","00:05:18.440","c3d8cf7","00:05:06.900","061e16d","00:05:03.000","eecc3cb","00:04:58.040",{"_220":1265,"_24":236,"_237":1266,"_239":1267},{"_220":1263,"_24":236,"_237":1264,"_239":279},{"_220":1260,"_24":236,"_237":1261,"_239":1262},{"_220":1258,"_24":236,"_237":1259,"_239":279},{"_220":1255,"_24":236,"_237":1256,"_239":1257},{"_220":1253,"_24":236,"_237":1254,"_239":279},{"_220":1250,"_24":236,"_237":1251,"_239":1252},"11a4950bba26",[1235],"it works across a lot of different models, and you are getting the outcomes that you actually want.","d43b01622e47",[],"714ff921bcb5",[1237],"for defining the outcome that you want your skills to have, and then it will optimize the text of the prompt in order for you to ensure that your prompt is using as few characters as possible,","59f956cd12ae",[],"114da9068125",[1239],"DSPy is a toolkit that you can use","fc897ba58f79",[],"2f937404e80f",[1241],"And another tip that I'll give you is you should look at DSPy to optimize your skills.","c8393bee2f0c",[1277,1278,1279],[1271,1272],{"_220":1275,"_24":231,"_232":1276},{"_220":1273,"_24":231,"_232":1274},"bbfdfe6","00:04:53.040","cdc325e","00:04:50.240",{"_220":1285,"_24":236,"_237":1286,"_239":1287},{"_220":1283,"_24":236,"_237":1284,"_239":279},{"_220":1280,"_24":236,"_237":1281,"_239":1282},"e48ee99d46e4",[1273],"It's really maturing. So some of the tools are here, you can take a look at them.","207a1333ee82",[],"c0e9f83bc152",[1275],"There is a lot of eval tooling that you should look into.","4f3c14e42974",[1294],[1291],{"_220":1292,"_24":231,"_232":1293},"78d0e89","00:04:43.040",{"_220":1295,"_24":236,"_237":1296,"_239":1297},"d18be905ace5",[1292],"If you have skills without evals, that is like code without unit tests, which means you cannot know that it's actually doing a better job over time.","71a01c792ab5",[1307,1308,1309],[1301,1302],{"_220":1305,"_24":231,"_232":1306},{"_220":1303,"_24":231,"_232":1304},"442945a","00:04:38.840","56de220","00:04:34.600",{"_220":1315,"_24":236,"_237":1316,"_239":1317},{"_220":1313,"_24":236,"_237":1314,"_239":279},{"_220":1310,"_24":236,"_237":1311,"_239":1312},"2858d6a2d92f",[1303],"And I want to stress to you that you must be writing evals for your skills, right?","f60652e4faf6",[],"12569b717708",[1305],"You add a phrase here, you add a phrase there, and the quality can super degrade.","3d424d5af04f",[1330,1331,1332,1333,1334],[1321,1322,1323],{"_220":1328,"_24":231,"_232":1329},{"_220":1326,"_24":231,"_232":1327},{"_220":1324,"_24":231,"_232":1325},"97ca1b2","00:04:22.620","36637af","00:04:19.220","3426b7f","00:04:12.160",{"_220":1345,"_24":236,"_237":1346,"_239":1347},{"_220":1343,"_24":236,"_237":1344,"_239":279},{"_220":1340,"_24":236,"_237":1341,"_239":1342},{"_220":1338,"_24":236,"_237":1339,"_239":279},{"_220":1335,"_24":236,"_237":1336,"_239":1337},"cc161065c5a3",[1324],"when you go to write a skill, I think the experience for most people is you wrote down your skill to say what you wanted to do and you have these magic words and it works the first time, but over time, the quality of your skill can degrade.","a19d995401f7",[],"779984004138",[1326],"And really,","fdc9d0790cf6",[],"5d89605ec5da",[1328],"So if memory is what your agents know, that's how we learn over time, skills are your agents' instructions.","4255350fe117",[1351],[],{"_220":1352,"_24":236,"_237":1353,"_239":1354},"8c37344c72f1",[],"Skills Need Evals","5b73bb4c3232",[1364,1365,1366],[1358,1359],{"_220":1362,"_24":231,"_232":1363},{"_220":1360,"_24":231,"_232":1361},"291e5ce","00:04:00.180","cfcc3c1","00:03:54.740",{"_220":1372,"_24":236,"_237":1373,"_239":1374},{"_220":1370,"_24":236,"_237":1371,"_239":279},{"_220":1367,"_24":236,"_237":1368,"_239":1369},"dc60427f13b0",[1360],"You should try it out. It's this sort of specification for LLM wikis and it's super good if you're looking for a really cheap file-based memory system for your agents. Check out open knowledge format.","9d1aa23d2206",[],"7ef54e058d27",[1362],"Another tip that I'll give you actually, just a hot tip, is if you haven't looked at Google's open knowledge format, it's very good.","64b629c729ee",[1387,1388,1389,1390,1391],[1378,1379,1380],{"_220":1385,"_24":231,"_232":1386},{"_220":1383,"_24":231,"_232":1384},{"_220":1381,"_24":231,"_232":1382},"0135534","00:03:44.700","dacdb77","00:03:41.340","fb5fcbd","00:03:34.840",{"_220":1402,"_24":236,"_237":1403,"_239":1404},{"_220":1400,"_24":236,"_237":1401,"_239":279},{"_220":1397,"_24":236,"_237":1398,"_239":1399},{"_220":1395,"_24":236,"_237":1396,"_239":279},{"_220":1392,"_24":236,"_237":1393,"_239":1394},"b0a05d498bb6",[1381],"streams like Slack and Google Docs and Asana, but essentially the traces that your agents have about what they tried to do, what they're working on, where they failed.","5d5becd769b4",[],"a7e63643ba1c",[1383],"Not only the context that lives within your","71d8bd62f5d9",[],"492e01f29c04",[1385],"And really it's super important if you're going to have a self-learning organization, maybe obvious, that you have to be aggregating your context.","8add5d8c6d21",[1414,1415,1416],[1408,1409],{"_220":1412,"_24":231,"_232":1413},{"_220":1410,"_24":231,"_232":1411},"2d8d29b","00:03:32.120","6a4cd26","00:03:22.380",{"_220":1422,"_24":236,"_237":1423,"_239":1424},{"_220":1420,"_24":236,"_237":1421,"_239":279},{"_220":1417,"_24":236,"_237":1418,"_239":1419},"47ade5e65d8d",[1410],"Your traces, task interactions, chat logs.","cef3458cf951",[],"d7379e688f54",[1412],"Memory is not just an API that you give to your agent to read a memory or write a memory, but essentially every piece of context that you have in your organization becomes effectively part of this memory, right?","8e33f10eaf97",[1440,1441,1442,1443,1444,1445,1446],[1428,1429,1430,1431],{"_220":1438,"_24":231,"_232":1439},{"_220":1436,"_24":231,"_232":1437},{"_220":1434,"_24":231,"_232":1435},{"_220":1432,"_24":231,"_232":1433},"7994183","00:03:18.860","93a3bb0","00:03:15.740","92d5cf6","00:03:11.659","9c0a529","00:03:07.180",{"_220":1462,"_24":236,"_237":1463,"_239":1464},{"_220":1460,"_24":236,"_237":1461,"_239":279},{"_220":1457,"_24":236,"_237":1458,"_239":1459},{"_220":1455,"_24":236,"_237":1456,"_239":279},{"_220":1452,"_24":236,"_237":1453,"_239":1454},{"_220":1450,"_24":236,"_237":1451,"_239":279},{"_220":1447,"_24":236,"_237":1448,"_239":1449},"b426cc7f0e46",[1432],"But I want to point out that memory is not just a product.","49bcfe7ea701",[],"7fe3db2a8eab",[1434],"We are also hearing from Letta later today.","910472985e67",[],"e1ec46d6ac5b",[1436],"We have some of these tools here, you can look.","d87337d313d8",[],"dd81a0f52c00",[1438],"So when we look at memory, there are a lot of tools that you can use for memory.","494b28da9415",[1477,1478,1479,1480,1481],[1468,1469,1470],{"_220":1475,"_24":231,"_232":1476},{"_220":1473,"_24":231,"_232":1474},{"_220":1471,"_24":231,"_232":1472},"aba1587","00:03:02.380","e670bee","00:02:57.180","f8a9c7d","00:02:50.940",{"_220":1492,"_24":236,"_237":1493,"_239":1494},{"_220":1490,"_24":236,"_237":1491,"_239":279},{"_220":1487,"_24":236,"_237":1488,"_239":1489},{"_220":1485,"_24":236,"_237":1486,"_239":279},{"_220":1482,"_24":236,"_237":1483,"_239":1484},"ec24c7ab05af",[1471],"What tools can your agent write? But we'll come back to that in a second.","b73b8353ca01",[],"a1e75fbedd13",[1473],"But essentially, when we give our agent tools, we immediately run into this problem, which is like, what tools can your agent read?","95d83ae6e1a5",[],"f93ab3ea4608",[1475],"And then you have the prompt for the task you have at hand, and essentially the world state, which is provided to you by the tools.","ddc1d0229427",[1507,1508,1509,1510,1511],[1498,1499,1500],{"_220":1505,"_24":231,"_232":1506},{"_220":1503,"_24":231,"_232":1504},{"_220":1501,"_24":231,"_232":1502},"d045812","00:02:46.960","c91e54f","00:02:38.540","f2dd8fe","00:02:30.860",{"_220":1522,"_24":236,"_237":1523,"_239":1524},{"_220":1520,"_24":236,"_237":1521,"_239":279},{"_220":1517,"_24":236,"_237":1518,"_239":1519},{"_220":1515,"_24":236,"_237":1516,"_239":279},{"_220":1512,"_24":236,"_237":1513,"_239":1514},"150cbc474671",[1501],"I mean, knowledge and memory are a bit fuzzy, but they are a bit different.","eeb760b61e4e",[],"cebe3e11b89c",[1503],"You have the agent's identity, you have knowledge about the world, you have memory. I'll define this as: memory has you and your tasks in it.","eae9771581ba",[],"0e2aa336bac7",[1505],"Your agent has an LLM which provides the intelligence, and then you have the context which you give to that intelligence, and context is not just one thing, right?","6156f859b0d0",[1534,1535,1536],[1528,1529],{"_220":1532,"_24":231,"_232":1533},{"_220":1530,"_24":231,"_232":1531},"64d566f","00:02:27.080","a872b9b","00:02:22.920",{"_220":1542,"_24":236,"_237":1543,"_239":1544},{"_220":1540,"_24":236,"_237":1541,"_239":279},{"_220":1537,"_24":236,"_237":1538,"_239":1539},"dc95f5496e13",[1530],"And so really when you dig into an agent, you want to look at these different pieces, right?","d31986959bb3",[],"1449e31b5d28",[1532],"Essentially they wake up with these superpowers, but they have no idea what they're supposed to do.","5b40f8b619fc",[1557,1558,1559,1560,1561],[1548,1549,1550],{"_220":1555,"_24":231,"_232":1556},{"_220":1553,"_24":231,"_232":1554},{"_220":1551,"_24":231,"_232":1552},"d49e260","00:02:21.480","4c9cdc8","00:02:16.040","bfe58b3","00:02:09.340",{"_220":1572,"_24":236,"_237":1573,"_239":1574},{"_220":1570,"_24":236,"_237":1571,"_239":279},{"_220":1567,"_24":236,"_237":1568,"_239":1569},{"_220":1565,"_24":236,"_237":1566,"_239":279},{"_220":1562,"_24":236,"_237":1563,"_239":1564},"5bf3b9c61b97",[1551],"And so that's what happens with your agent.","f0da1a0ab5d0",[],"0c5c32ab9b85",[1553],"who he is, where he's going, what he's supposed to do, and so he draws tattoos on his arms and takes Polaroids.","de29a85877ca",[],"62252d4be696",[1555],"This guy, he wakes up, he finds himself in a hotel room and he's got a gun and he's really capable and could drive a car, but he has no idea","031d1fca0b94",[1584,1585,1586],[1578,1579],{"_220":1582,"_24":231,"_232":1583},{"_220":1580,"_24":231,"_232":1581},"e0bbfff","00:02:06.220","0a33b93","00:02:04.380",{"_220":1592,"_24":236,"_237":1593,"_239":1594},{"_220":1590,"_24":236,"_237":1591,"_239":279},{"_220":1587,"_24":236,"_237":1588,"_239":1589},"a20138c7e66c",[1580],"Yeah, exactly. And essentially your agents are like Memento man.","50a7f13ac50c",[],"a76be9949cca",[1582],"It's a 2000s Christopher Nolan movie.","e421e81b714a",[1604,1605,1606],[1598,1599],{"_220":1602,"_24":231,"_232":1603},{"_220":1600,"_24":231,"_232":1601},"ac5b857","00:01:59.540","1c7a2b9","00:01:53.619",{"_220":1612,"_24":236,"_237":1613,"_239":1614},{"_220":1610,"_24":236,"_237":1611,"_239":279},{"_220":1607,"_24":236,"_237":1608,"_239":1609},"8cfbaba71f29",[1600],"Like, how does an agent actually learn? I don't know if any of you have seen the movie Memento.","f1f308477d73",[],"de5b2eb66611",[1602],"And so really if we're gonna have an organization that learns, that really starts with looking at the agent themselves.","5f334badc69f",[1618],[],{"_220":1619,"_24":236,"_237":1620,"_239":1621},"036e6a2091d1",[],"Inside the Agent","c0bff18df638",[1631,1632,1633],[1625,1626],{"_220":1629,"_24":231,"_232":1630},{"_220":1627,"_24":231,"_232":1628},"8d74293","00:01:48.340","1d9133c","00:01:43.619",{"_220":1639,"_24":236,"_237":1640,"_239":1641},{"_220":1637,"_24":236,"_237":1638,"_239":279},{"_220":1634,"_24":236,"_237":1635,"_239":1636},"c520e8f62d8a",[1627],"How do you define insight? You have these software tasks, how do you define what's good enough for you?","6ae68ad6d969",[],"82b74d09bf70",[1629],"But when we come to other tasks other than code, you have this question like, how do you define your taste?","ee4dfbc482d2",[1651,1652,1653],[1645,1646],{"_220":1649,"_24":231,"_232":1650},{"_220":1647,"_24":231,"_232":1648},"f2bc7a3","00:01:39.760","83226ec","00:01:29.840",{"_220":1659,"_24":236,"_237":1660,"_239":1661},{"_220":1657,"_24":236,"_237":1658,"_239":279},{"_220":1654,"_24":236,"_237":1655,"_239":1656},"97cb5d12d4d1",[1647],"Because code has this provable outcome that you can define.","5f8c0d53d6ab",[],"78d5524f886c",[1649],"And we've seen this with code, actually, is the first really task to fall to the AI working well, right?","02f6d0375f37",[1671,1672,1673],[1665,1666],{"_220":1669,"_24":231,"_232":1670},{"_220":1667,"_24":231,"_232":1668},"c3c37a5","00:01:24.320","83befaf","00:01:15.759",{"_220":1679,"_24":236,"_237":1680,"_239":1681},{"_220":1677,"_24":236,"_237":1678,"_239":279},{"_220":1674,"_24":236,"_237":1675,"_239":1676},"b22d6640b994",[1667],"And that's really a result of not telling the AI what you actually want it to do.","3262b06543eb",[],"faa7ac95772b",[1669],"This famously is the paperclip problem, the idea that an AI will be misaligned and it will compete with us for resources and essentially have a poor outcome.","d5f48445435a",[1691,1692,1693],[1685,1686],{"_220":1689,"_24":231,"_232":1690},{"_220":1687,"_24":231,"_232":1688},"c5584a0","00:01:10.200","deb4a03","00:01:07.080",{"_220":1699,"_24":236,"_237":1700,"_239":1701},{"_220":1697,"_24":236,"_237":1698,"_239":279},{"_220":1694,"_24":236,"_237":1695,"_239":1696},"184924d43dfe",[1687],"you can improve your system if it's self-learning, and if it's autonomous, it can improve itself in the wrong direction.","c315dab8566c",[],"4eb8ea89e8b1",[1689],"So what's the problem? One of the biggest problems that you can have:","c477c970df14",[1714,1715,1716,1717,1718],[1705,1706,1707],{"_220":1712,"_24":231,"_232":1713},{"_220":1710,"_24":231,"_232":1711},{"_220":1708,"_24":231,"_232":1709},"aa8cb7a","00:01:00.320","032a000","00:00:56.960","7d8498a","00:00:52.640",{"_220":1729,"_24":236,"_237":1730,"_239":1731},{"_220":1727,"_24":236,"_237":1728,"_239":279},{"_220":1724,"_24":236,"_237":1725,"_239":1726},{"_220":1722,"_24":236,"_237":1723,"_239":279},{"_220":1719,"_24":236,"_237":1720,"_239":1721},"21a49013bdaa",[1708],"So by the end of this talk I'm going to give you eight layers of the organization that improves itself.","bfcb56dd7a8d",[],"cb40b53e93a9",[1710],"It's open source. You can download it today and try it out.","c217af9ec5f4",[],"d15c9787be6a",[1712],"So it is a toolkit that you use for humans and agents to work together.","899d3ef04eca",[1741,1742,1743],[1735,1736],{"_220":1739,"_24":231,"_232":1740},{"_220":1737,"_24":231,"_232":1738},"30d01ac","00:00:49.280","72613ac","00:00:46.120",{"_220":1749,"_24":236,"_237":1750,"_239":1751},{"_220":1747,"_24":236,"_237":1748,"_239":279},{"_220":1744,"_24":236,"_237":1745,"_239":1746},"7a46f1ae4365",[1737],"We have 80,000 GitHub stars. It is a meta harness.","9c4689fcbed0",[],"ca9f859c1880",[1739],"So I, yeah, the creator of Paperclip.","e0ae83366c50",[1764,1765,1766,1767,1768],[1755,1756,1757],{"_220":1762,"_24":231,"_232":1763},{"_220":1760,"_24":231,"_232":1761},{"_220":1758,"_24":231,"_232":1759},"d0518a4","00:00:42.920","1666e0f","00:00:39.880","623ad48","00:00:37.880",{"_220":1779,"_24":236,"_237":1780,"_239":1781},{"_220":1777,"_24":236,"_237":1778,"_239":279},{"_220":1774,"_24":236,"_237":1775,"_239":1776},{"_220":1772,"_24":236,"_237":1773,"_239":279},{"_220":1769,"_24":236,"_237":1770,"_239":1771},"a74b4670449a",[1758],"And you need to be able to improve that and feed that back into the system.","8efe1a2b3b2f",[],"1f3ac62ee811",[1760],"Can you judge if what they're doing is actually good or helpful to you?","55d1ed9bb6a2",[],"8a7873b787ee",[1762],"You need to be able to evaluate their outcomes.","5c1ad698db08",[1797,1798,1799,1800,1801,1802,1803],[1785,1786,1787,1788],{"_220":1795,"_24":231,"_232":1796},{"_220":1793,"_24":231,"_232":1794},{"_220":1791,"_24":231,"_232":1792},{"_220":1789,"_24":231,"_232":1790},"3f79f65","00:00:34.920","3c225cb","00:00:32.279","13e38d7","00:00:29.800","26ea190","00:00:24.400",{"_220":1819,"_24":236,"_237":1820,"_239":1821},{"_220":1817,"_24":236,"_237":1818,"_239":279},{"_220":1814,"_24":236,"_237":1815,"_239":1816},{"_220":1812,"_24":236,"_237":1813,"_239":279},{"_220":1809,"_24":236,"_237":1810,"_239":1811},{"_220":1807,"_24":236,"_237":1808,"_239":279},{"_220":1804,"_24":236,"_237":1805,"_239":1806},"177954d81894",[1789],"You need to be able to control what they're doing.","67c35d1bd83a",[],"0de596d369bc",[1791],"You need to be able to control what they have access to.","f083d4e6a2d9",[],"ab5973840745",[1793],"You need to understand what your agents are doing.","3a996542c78d",[],"73d2b2ef65c6",[1795],"And really when it comes to working with agents, I think you really have this four-step process.","5b337c99720e",[1831,1832,1833],[1825,1826],{"_220":1829,"_24":231,"_232":1830},{"_220":1827,"_24":231,"_232":1828},"b12eb05","00:00:18.180","ab96727","00:00:13.780",{"_220":1839,"_24":236,"_237":1840,"_239":1841},{"_220":1837,"_24":236,"_237":1838,"_239":279},{"_220":1834,"_24":236,"_237":1835,"_239":1836},"63b1bf9107b4",[1827],"So you have this action, you measure that action, and then you update based on the results of what you did.","c3ecaee1dc75",[],"3d0ebee0cec1",[1829],"And I think one of the things you should look at is that everything that learns, it really runs the same loop.","6641999a9c32",[1851,1852,1853],[1845,1846],{"_220":1849,"_24":231,"_232":1850},{"_220":1847,"_24":231,"_232":1848},"7c2f874","00:00:08.400","cbdea2b","00:00:03.040",{"_220":1859,"_24":236,"_237":1860,"_239":1861},{"_220":1857,"_24":236,"_237":1858,"_239":279},{"_220":1854,"_24":236,"_237":1855,"_239":1856},"f7a797dfc99a",[1847],"So the question we have here is, how do we create autonomous organizations that learn?","a624ce9e76d7",[],"59eb599333b5",[1849],"My name is Dotta and I'm gonna talk about the self-improving organization.","0d480443dee7",[1865],[],{"_220":1866,"_24":236,"_237":1867,"_239":1868},"db9467c77193",[],"The Self-Improving Organization",{"_18":1886,"_20":1911,"_22":1888,"_24":1878,"_26":1889,"_1879":1912,"_1881":1913,"_14":1914},{"_18":1886,"_20":1906,"_22":1888,"_24":1878,"_26":1889,"_1879":1907,"_1881":1908,"_14":1909},{"_18":1899,"_20":1900,"_22":1901,"_24":1878,"_26":1899,"_1879":1902,"_1881":1903,"_14":1904},{"_18":1886,"_20":1894,"_22":1888,"_24":1878,"_26":1889,"_1879":1895,"_1881":1896,"_14":1897},{"_18":1886,"_20":1887,"_22":1888,"_24":1878,"_26":1889,"_1879":1890,"_1881":1891,"_14":1892},{"_18":1875,"_20":1876,"_22":1877,"_24":1878,"_26":1875,"_1879":1880,"_1881":1882,"_14":1883},"2023-07-17T23:23:49Z","eec559a9-10e7-4405-82e4-e0736495ec36","TfMkf5gVJ7UWv1vMi5Qxor","tag","description","Artificial intelligence, including AI-powered coding assistants built on large language models.","name","AI",{"_24":14,"_1884":1885},"current","ai","2026-10-23T23:10:26Z","d85bfd87-b157-4c36-9e8f-16f6cfb6ac69","mG7vmCjaB1vEw3gqUIqrjf","2026-10-23T23:11:12Z","Keeping autonomous agents and organizations safe and aligned.","AI Safety",{"_24":14,"_1884":1893},"ai-safety","8d2828f0-70fa-49f2-a7cf-4b648e941299","Coordinating many agents, goals, and roles.","Agent Orchestration",{"_24":14,"_1884":1898},"agent-orchestration","2025-04-15T17:19:22Z","80867d7d-9222-41f5-a465-915aa791bc97","YlxgmKTLxstUdop23VNy4f","AI agents are autonomous systems built to independently take action on business or technical tasks.","AI Agents",{"_24":14,"_1884":1905},"ai-agents","6625b640-07c5-4ac4-9324-bf13e52c5ed3","Agents that work toward goals over hours or days, and the infrastructure that keeps them on track.","Long-Horizon Agents",{"_24":14,"_1884":1910},"long-horizon-agents","55e6b0bf-3cf3-4fb8-b73d-a89a18180b76","Evaluating agent behavior, skills, and learning.","Agent Evals",{"_24":14,"_1884":1915},"agent-evals",{"_220":1917,"_24":1918,"_1919":-5,"_1920":1921},"spk0dottab","speaker","organization","person",{"_18":1922,"_20":1923,"_22":1924,"_1925":1926,"_24":1920,"_26":1927,"_1928":1929,"_1930":12,"_1881":1931,"_1919":1932,"_1933":12,"_1934":1935,"_1936":1937,"_68":1938,"_1939":12},"2026-09-10T17:40:49Z","ad64b1f2-11e4-4aa6-9270-a43f6c836ba2","AvBrD6DdjyccEYuoF7sQdC","_system",{"_1969":1970},"2026-09-10T18:13:35Z","bio",[1953],"generate","Dotta B",{"_18":1947,"_20":1948,"_22":1949,"_24":1919,"_26":1950,"_1930":12,"_1881":1951,"_1952":12},"partner","photo",{"_24":1942,"_74":1943},"position","Cofounder",{"_24":68,"_1940":12,"_1941":12},"teamMember","nofollow","noindex","mainImage",{"_1944":1945,"_24":1946},"_ref","image-7e9f992a5761a6638b2e5b7718919d8699455d7c-682x682-jpg","reference","2026-09-10T17:40:01Z","b34d14d7-6422-4e07-9da3-b5e465c34a1e","AvBrD6DdjyccEYuoF7qYYQ","2026-09-10T17:40:13Z","Paperclip","portfolioCompany",{"_220":1954,"_24":222,"_223":1955,"_225":1956,"_227":228},"b1",[1962,1963],[1957],{"_220":1958,"_24":1959,"_1960":1961},"l1","link","href","https://www.linkedin.com/in/natemurray/",{"_220":1967,"_24":236,"_237":1968,"_239":1931},{"_220":1964,"_24":236,"_237":1965,"_239":1966},"s2",[]," is Cofounder of Paperclip, where he works on building agent organizations around measurable value: how agents decompose goals, route work, and negotiate capabilities with one another.","s1",[1958],"base",{"_100":1923,"_1971":1972},"rev","IBgcBIPGoUVrOhZYuqyn3s","image",{"_1944":1975,"_24":1946},"image-e10d4d928c257f1665bbe8d6506f20d720fbb570-1920x1080-jpg",{"_220":1977,"_24":222,"_223":1978,"_225":1979,"_227":228},"be3d603371f3",[1980],[],{"_220":1981,"_24":236,"_237":1982,"_239":1983},"d49ccbbfd263",[],"

Goals! Orchestration and the Human Control Plane\n
- The Self-Improving Organization
- Inside the Agent
- Skills Need Evals
- Trust and Agent Roles
- Goals and the Eight Layers
\n

",{"_220":1991,"_1944":1992,"_24":1946},{"_220":1989,"_1944":1990,"_24":1946},{"_220":1987,"_1944":1988,"_24":1946},"1gQ61dhASaRl","67355535-cdb5-472f-b416-5b48b1294f09","BBJ5gEVyR649","4cf0dd20-6e50-4e54-91f2-18e60d34af37","gTzWBWaKwlkD","27ad23dc-58b9-4366-a3b9-c73d40aedc73","items",[4455,4456,4457],"recent",[1997,1998,1999],{"_2000":3837,"_96":3838,"_1973":-5,"_2004":3839,"_51":3840,"_14":3841,"_2008":2009,"_63":3842,"_2011":2012},{"_2000":2617,"_96":2618,"_1973":-5,"_2004":2619,"_51":2620,"_14":2621,"_2008":2009,"_63":2622,"_2011":2012},{"_2000":2001,"_96":2002,"_1973":2003,"_2004":2005,"_51":2006,"_14":2007,"_2008":2009,"_63":2010,"_2011":2012},"content",[2013,2014,2015,2016,2017,2018,2019,2020,2021,2022,2023,2024,2025,2026,2027,2028,2029,2030,2031,2032,2033,2034,2035,2036,2037,2038,2039,2040,2041,2042,2043,2044,2045,2046,2047,2048,2049,2050,2051,2052,2053,2054,2055,2056,2057],17,"https://cdn.sanity.io/images/50q6fr1p/production/28801f05c737cc57c3c04d467a0a54838fc70ad0-1200x630.png","objectID","9d8377ac-25e6-47b7-8e40-22ee9a7cb43e","2026-08-27T16:38:00.000Z","understanding-code-by-treating-it-as-a-graph","templateType","default","Is the Key to Understanding Code Treating it as a Graph?","typeName","blogArticle",{"_220":2610,"_24":222,"_223":2611,"_225":2612,"_227":576},{"_220":2581,"_24":222,"_223":2582,"_225":2583,"_227":228},{"_220":2552,"_24":222,"_223":2553,"_225":2554,"_227":228},{"_220":2549,"_24":1973,"_74":2550},{"_220":2536,"_24":222,"_223":2537,"_225":2538,"_227":228},{"_220":2529,"_24":222,"_223":2530,"_225":2531,"_227":576},{"_220":2511,"_24":222,"_223":2512,"_225":2513,"_227":228},{"_220":2493,"_24":222,"_223":2494,"_225":2495,"_227":228},{"_220":2486,"_24":222,"_223":2487,"_225":2488,"_227":228},{"_220":2479,"_24":222,"_223":2480,"_225":2481,"_227":228},{"_220":2472,"_24":222,"_223":2473,"_225":2474,"_227":576},{"_220":2443,"_24":222,"_223":2444,"_225":2445,"_227":228},{"_220":2436,"_24":222,"_223":2437,"_225":2438,"_227":228},{"_220":2429,"_24":222,"_223":2430,"_225":2431,"_227":228},{"_220":2422,"_24":222,"_223":2423,"_225":2424,"_227":576},{"_220":2415,"_24":222,"_223":2416,"_225":2417,"_227":228},{"_220":2397,"_24":222,"_223":2398,"_225":2399,"_227":228},{"_220":2379,"_24":222,"_223":2380,"_225":2381,"_227":228},{"_220":2372,"_24":222,"_223":2373,"_225":2374,"_227":576},{"_220":2365,"_24":222,"_223":2366,"_225":2367,"_227":228},{"_220":2339,"_24":222,"_223":2340,"_225":2341,"_227":228},{"_220":2321,"_24":222,"_223":2322,"_225":2323,"_227":228},{"_220":2306,"_24":222,"_223":2307,"_225":2308,"_227":228},{"_220":2299,"_24":222,"_223":2300,"_225":2301,"_227":228},{"_220":2292,"_24":222,"_223":2293,"_225":2294,"_227":228},{"_220":2285,"_24":222,"_223":2286,"_225":2287,"_227":576},{"_220":2278,"_24":222,"_223":2279,"_225":2280,"_227":228},{"_220":2271,"_24":222,"_223":2272,"_225":2273,"_227":228},{"_220":2264,"_24":222,"_223":2265,"_225":2266,"_227":228},{"_220":2235,"_24":222,"_223":2236,"_225":2237,"_227":228},{"_220":2228,"_24":222,"_223":2229,"_225":2230,"_227":576},{"_220":2221,"_24":222,"_223":2222,"_225":2223,"_227":228},{"_220":2214,"_24":222,"_223":2215,"_225":2216,"_227":228},{"_220":2207,"_24":222,"_223":2208,"_225":2209,"_227":228},{"_220":2200,"_24":222,"_223":2201,"_225":2202,"_227":228},{"_220":2193,"_24":222,"_223":2194,"_225":2195,"_227":576},{"_220":2186,"_24":222,"_223":2187,"_225":2188,"_227":228},{"_220":2146,"_24":222,"_223":2147,"_225":2148,"_227":228},{"_220":2117,"_24":222,"_223":2118,"_225":2119,"_227":228},{"_220":2110,"_24":222,"_223":2111,"_225":2112,"_227":576},{"_220":2103,"_24":222,"_223":2104,"_225":2105,"_227":228},{"_220":2088,"_24":222,"_223":2089,"_225":2090,"_227":228},{"_220":2081,"_24":222,"_223":2082,"_225":2083,"_227":228},{"_220":2065,"_24":222,"_223":2066,"_225":2067,"_227":228},{"_220":2058,"_24":222,"_223":2059,"_225":2060,"_227":228},"8eaf5dda166e",[2061],[],{"_220":2062,"_24":236,"_237":2063,"_239":2064},"7fccb285f275",[],"“So I think orchestration will be important for quite some time, because AI is still not at the level of fully, autonomously orchestrating large-scale systems that will work smoothly and without any problems. We are not there yet.”","628421de9da8",[2068,2069,2070],[],{"_220":2078,"_24":236,"_237":2079,"_239":2080},{"_220":2074,"_24":236,"_237":2075,"_239":2076},{"_220":2071,"_24":236,"_237":2072,"_239":2073},"ad91e713dbfb",[],", but also understanding what caching is and why it’s important. You’re going to need to apply that understanding when you think your system actually needs caching!”","d4b8c4a82cf1",[2077],"what Redis is","em","0cc52c592750",[],"“It’s about having this type of ‘systems thinking,’ an understanding of system design architecture and the underlying principles of the technologies. Not just knowing ","dc75d7ad8097",[2084],[],{"_220":2085,"_24":236,"_237":2086,"_239":2087},"c6a9980c57d2",[],"“You need to understand how the distributed systems work, and the code-graph-rag project fits really well here to help you ask architectural questions and locate potential bottlenecks, like dead code or other opportunities for optimization.”","afb3c879f969",[2091,2092,2093],[],{"_220":2100,"_24":236,"_237":2101,"_239":2102},{"_220":2097,"_24":236,"_237":2098,"_239":2099},{"_220":2094,"_24":236,"_237":2095,"_239":2096},"56e7f92d1742",[]," to use it, and what the pros and cons are. That will help you to have this overarching, holistic view of what it's like to build a highly scalable software engineering system.”","7f5989d240bd",[2077],"when","49750bd53c8c",[],"“For example, in order to deploy applications at scale, you might need a Kubernetes cluster. You shouldn’t have to be a highly skilled DevOps engineer to know that as AI can help you there, but you would need to know ","ab8bfa185ba6",[2106],[],{"_220":2107,"_24":236,"_237":2108,"_239":2109},"067bce70e73c",[],"From his experience building the project, Avagyan advises anyone working in a technical capacity to learn how to orchestrate. “Software engineering is changing faster than people think. The ability to orchestrate things will be important no matter what. By that, I mean knowing many different technologies: Their strengths and weaknesses, the scalability of each thing, and when to use what.”","1d6c6865362e",[2113],[],{"_220":2114,"_24":236,"_237":2115,"_239":2116},"2586e6d9e8b9",[],"The Future of Codebases: Built and Managed by AI?","6869a525750a",[2126,2127,2128,2129,2130],[2120,2121],{"_220":2124,"_24":1959,"_1960":2125},{"_220":2122,"_24":1959,"_1960":2123},"41760ee29615","https://www.youtube.com/watch?v=O2jUTq6nCEY&t=2s","48182a105088","https://blog.dailydoseofds.com/",{"_220":2143,"_24":236,"_237":2144,"_239":2145},{"_220":2140,"_24":236,"_237":2141,"_239":2142},{"_220":2137,"_24":236,"_237":2138,"_239":2139},{"_220":2134,"_24":236,"_237":2135,"_239":2136},{"_220":2131,"_24":236,"_237":2132,"_239":2133},"dd3d4a61f74f",[],"), but eventually, the project started trending again on GitHub organically. I guess it might have to do with the ‘right solution at the right time’ phenomenon.”","6f0d56991c70",[2122],"with Memgraph","3954a9df1f48",[]," when I was working on my own startup back in 2023. I got coverage for my product in this popular newsletter, which might have been the trigger that helped it ‘go viral’ on GitHub the first time. That said, I have also tried many other channels like Hacker News, LinkedIn and X posts, giving talks (including this session ","6424a1b96d65",[2124],"Daily Dose of Data Science","30637c109bb6",[],"“Finally, I’d recommend finding a distribution channel for your work. Even the greatest product on earth won’t necessarily be noticed if you just put it on GitHub and leave it there. You need to promote it. I was fortunate enough to have worked with the founder of ","81d7413fc3b2",[2158,2159,2160,2161,2162,2163,2164],[2149,2150,2151],{"_220":2156,"_24":1959,"_1960":2157},{"_220":2154,"_24":1959,"_1960":2155},{"_220":2152,"_24":1959,"_1960":2153},"2f238ed440f7","https://zed.dev/","3f971d287b17","https://neovim.io/","f040745fb71d","https://github.com/vitali87/croft",{"_220":2183,"_24":236,"_237":2184,"_239":2185},{"_220":2180,"_24":236,"_237":2181,"_239":2182},{"_220":2177,"_24":236,"_237":2178,"_239":2179},{"_220":2174,"_24":236,"_237":2175,"_239":2176},{"_220":2171,"_24":236,"_237":2172,"_239":2173},{"_220":2168,"_24":236,"_237":2169,"_239":2170},{"_220":2165,"_24":236,"_237":2166,"_239":2167},"1bb94c846e83",[]," and other alternatives. I found myself thinking there should be a better way, so I built a VS Code clone in the terminal with Rust to make it super-fast. I’ve been using it every day myself and I would recommend others try it out.”","93a91dcf4857",[2152],"Zed","e76bcfb5a8f3",[],", ","94d7ac6fa32a",[2154],"Neovim","630e70cbbe50",[],", as an example. “I’m a VS Code user, but I enjoy terminal as well, so I found myself going back and forth between VS Code, ","9f8ed6fdab01",[2156],"Croft","321ac6bc0e40",[],"“Second, I would recommend building things from the perspective of thinking about what you do every day and how you can accelerate what you do and be more performant with your tools.” The creator notes another open-source project of his, ","8c4981547047",[2189],[],{"_220":2190,"_24":236,"_237":2191,"_239":2192},"58335d9df518",[],"When asked what his most important learnings from the project are, Avagyan offers a few suggestions. “One, which I think is most important, is to just build something for yourself that you would like to have. That's pretty much how this project was born. Because I was dealing with [the challenges of managing a] monorepo, and at first, I just wanted to visualize it and understand its complexity. I looked for solutions to see if there was a good product for me. I didn't find one, so I decided I would build it myself, and after releasing it publicly, it turned to be useful for other people.”","05e8e1c9fe0a",[2196],[],{"_220":2197,"_24":236,"_237":2198,"_239":2199},"76872200c049",[],"Learnings from Building","526da0bc67f6",[2203],[],{"_220":2204,"_24":236,"_237":2205,"_239":2206},"b5c434461aac",[],"The creator suggests that while he spent the better part of a year building a system that could dynamically map and remap graph representations of codebases without significant latency, not every org is on the same page. “Even doing something like this with AI agents will take some time, and will cost internal resources. And people don’t always have patience for these things and may want to see a one-shot solution instead. But I think the approach will get more momentum and I think many companies will adopt it eventually.”","f2f91f55c9c2",[2210],[],{"_220":2211,"_24":236,"_237":2212,"_239":2213},"5556895009c6",[],"“You can get really far with one language, maybe with a toy example using a tool like Tree-sitter. But jobs that require you to unify graphs across different languages, and build usable and accurate relationships over top of them, and making sure it can work at scale, that’s different. For example, a monorepo is a living creature that changes every minute: People commit and delete code and so on. So your graph needs to keep up with that. It needs to recalculate [those relationships] with every change, every drop, every addition. The project has a graph updater designed specifically to make this real-time syncing possible.”","a89686ddcbd7",[2217],[],{"_220":2218,"_24":236,"_237":2219,"_239":2220},"eb00873aac4a",[],"Avagyan notes that those companies may be hitting a stumbling block. “I think one problem companies may be having is accuracy.”","7cbaec07cf9b",[2224],[],{"_220":2225,"_24":236,"_237":2226,"_239":2227},"52c6bc34d2b2",[],"The creator suggests that graphing codebases is a useful approach for any organization, but the timing might not be quite right for large-scale enterprises that are slow to embrace change. “I believe many companies, including major AI vendors like Cursor or Anthropic, have been experimenting with graph-based approaches, if not for codebase mapping, then for use cases like memory management.”","5a3ea8acd83b",[2231],[],{"_220":2232,"_24":236,"_237":2233,"_239":2234},"1cd77c6dd0f6",[],"Why Aren’t Enterprises Already Graphing Their Codebases?","b003dd295bbf",[2244,2245,2246,2247,2248],[2238,2239],{"_220":2242,"_24":1959,"_1960":2243},{"_220":2240,"_24":1959,"_1960":2241},"eb5b74dd23fd","https://flutter.dev/","fc884ca87ef9","https://dart.dev/overview",{"_220":2261,"_24":236,"_237":2262,"_239":2263},{"_220":2258,"_24":236,"_237":2259,"_239":2260},{"_220":2255,"_24":236,"_237":2256,"_239":2257},{"_220":2252,"_24":236,"_237":2253,"_239":2254},{"_220":2249,"_24":236,"_237":2250,"_239":2251},"b630193673be",[]," widget structure.”","ab660c7a1ffd",[2240],"Flutter","cee0616ee44a",[]," is already supported and includes ","fc85ec3d18b6",[2242],"Dart","2a8e4e8ba899",[],"Avagyan notes growing interest among teams managing heavyweight ERP systems with some flavor of Python, as well as ongoing requests around app development. “","e96ed7f2d62e",[2267],[],{"_220":2268,"_24":236,"_237":2269,"_239":2270},"e4b2c56f631d",[],"So you can see why companies are adopting the project as part of their onboarding tools. You could use it as an MCP server in Claude Code or any other harness or just use Cypher queries to search your codebase before even starting any agentic exploration. This approach is quick and gives you a really good grasp of a codebase, without getting sidetracked to other areas that may not be as important.”","a9bdfd671d83",[2274],[],{"_220":2275,"_24":236,"_237":2276,"_239":2277},"b0569066d73c",[],"“Since we recently added FLOWS_TO, READS_FROM and WRITES_TO relationships in a graph, it’s now easy to analyze the relationships between microservices in your codebase as well. (I’ve actually been approached by security companies about using the project for security detection use cases across large codebases, which is a natural fit as tracing how data moves through a system is exactly what those relationships capture.)”","34664692f09c",[2281],[],{"_220":2282,"_24":236,"_237":2283,"_239":2284},"2b1eaf954306",[],"The creator calls out onboarding new engineers as a key use case for the project. “People tell me that using the project has significantly reduced onboarding times because it can gather information about different aspects of a monorepo codebase and give that directly to new junior engineers.”","06fc03e6d208",[2288],[],{"_220":2289,"_24":236,"_237":2290,"_239":2291},"0743e3b412d6",[],"Real-World Use Cases for Graphing a Codebase","6a4ddfd50139",[2295],[],{"_220":2296,"_24":236,"_237":2297,"_239":2298},"847581b898f9",[],"“To be honest, the goal of the project was never cost reduction, but I think it’s a byproduct, since you get the accuracy of a graph and the accuracy of an MCP and agent to invoke the correct tools to query the graph.","2cfb4c4a9a81",[2302],[],{"_220":2303,"_24":236,"_237":2304,"_239":2305},"b8c18965ccb1",[],"“Let’s say the project fails in one Cypher query? Then it can just repeat it, and with that query, you can gain lots of context up front, very fast. And that failure is less costly compared to if AI were to go and perform its own search, realize it was searching in the wrong place, then go to another place...you can imagine how things would accumulate really fast.”","824172ebddac",[2309,2310,2311],[],{"_220":2318,"_24":236,"_237":2319,"_239":2320},{"_220":2315,"_24":236,"_237":2316,"_239":2317},{"_220":2312,"_24":236,"_237":2313,"_239":2314},"e051511f75df",[]," say that I’ve observed the project to be very fast, and it seems more fault-tolerant in a way.”","36ec44554375",[2077],"can","7cdd05898655",[],"Avagyan notes that some users come to the project looking for performance and cost gains as well. “Some people claim that the project is as much as 65% more cost-efficient than traditional agentic coding and search, and I can’t verify those numbers myself. I ","623b0301e0bf",[2327,2328,2329],[2324],{"_220":2325,"_24":1959,"_1960":2326},"c1ed526f5769","https://memgraph.com/",{"_220":2336,"_24":236,"_237":2337,"_239":2338},{"_220":2333,"_24":236,"_237":2334,"_239":2335},{"_220":2330,"_24":236,"_237":2331,"_239":2332},"da89921f17d9",[]," as a GraphDB and it's really fast, it runs in memory and you can deploy using Docker containers, and that would be enough to deploy locally, and the project has the building blocks for bigger deployments to the cloud. So far, it seems to scale well (people have even graphed the Linux kernel) and I haven’t seen any real degradation in speed when it comes to querying graphs.”","34da333815d3",[2325],"Memgraph","c9fdbc40e4d6",[],"“Otherwise, I don't think storage is a problem. The project uses ","d8236069c0ba",[2345,2346,2347,2348,2349],[2342],{"_220":2343,"_24":1959,"_1960":2344},"ec14de8b6db9","https://en.wikipedia.org/wiki/Dead-code_elimination",{"_220":2362,"_24":236,"_237":2363,"_239":2364},{"_220":2359,"_24":236,"_237":2360,"_239":2361},{"_220":2356,"_24":236,"_237":2357,"_239":2358},{"_220":2353,"_24":236,"_237":2354,"_239":2355},{"_220":2350,"_24":236,"_237":2351,"_239":2352},"2dd3e29f6f5e",[]," command that will check the reachability of all your code. If some are not reachable, they’re likely candidates [for being dead code and potentially getting deleted].”","cba21817906a",[2077],"code graph","28ed3f1a3a15",[],". If you have vibe coded your repo, or it’s otherwise in a state that you don't know exactly what's happening because there are so many abstractions or code duplications, or that code isn’t reachable. For example, there is a ","0f95656bd1a7",[2343],"dead code detection","37e4c8419854",[],"“I would say that a practical use of code-graph-rag is ","2e27e83e6ea0",[2368],[],{"_220":2369,"_24":236,"_237":2370,"_239":2371},"b4e8f24bb41b",[],"The creator suggests that the project is continuing to improve on traditional challenges in graph-based projects, which have historically not always been particularly responsive. “I would say storage has never been a problem. With this project, I spent significant time optimizing the build process of going from codebase to graph. (I actually optimised the build time by more than 5x with a separate repo I built that optimizes user-defined codebase segments against specific metrics like speed within accepted memory constraints. The agent then \"autosearches\" the best code among competing candidates while running the build against the ground-truthed evaluation cases which you or your agents have collected or constructed yourself.)”","e71e561bda30",[2375],[],{"_220":2376,"_24":236,"_237":2377,"_239":2378},"8bfe3199c316",[],"Bottlenecks and Resource Management","9b53b53a2ef7",[2385,2386,2387],[2382],{"_220":2383,"_24":1959,"_1960":2384},"cb676615c060","https://simonwillison.net/guides/agentic-engineering-patterns/red-green-tdd/",{"_220":2394,"_24":236,"_237":2395,"_239":2396},{"_220":2391,"_24":236,"_237":2392,"_239":2393},{"_220":2388,"_24":236,"_237":2389,"_239":2390},"a3e38bb4ef42",[],". Over time, the combination of hand-built code and agent-powered testing and development has proven to be effective. “At the beginning, I spent so much time on architecting [the project], setting the right relationships, that sort of thing, to ensure the foundation was ‘right.’ But when the foundation is ‘right,’ AI is great at doing the rest.”","a90a0f3ba6d2",[2383],"red/green test-driven development","84140f1e533e",[],"“And from there, we run thousands of test cases to ensure there are no regressions every time a new functionality is added. Every new functionality or bugfix is accompanied by a handful of new tests.” The creator suggests the project continuously runs more than 7,000 unit/integration tests using ","4ed1d30f3dc2",[2403,2404,2405],[2400],{"_220":2401,"_24":1959,"_1960":2402},"1b1e54e0707e","https://en.wikipedia.org/wiki/Polymorphism_(computer_science)",{"_220":2412,"_24":236,"_237":2413,"_239":2414},{"_220":2409,"_24":236,"_237":2410,"_239":2411},{"_220":2406,"_24":236,"_237":2407,"_239":2408},"9b109a84fb06",[]," works.” The creator notes that tracking all the nuances isn’t necessarily covered by existing deterministic open-source tools either. What is even more difficult is the runtime behavior of the code. That's why recent efforts have been placed to trace call relationships apparent at runtime and merge them into the same graph with statically-derived ones.”","e56a902ff095",[2401],"polymorphism","9a4ee9a258b2",[],"“Obviously, there are so many considerations when you go into specifics, like finding call relationships. Each language has its own specific intricacies, like how ","0815be5d9aee",[2418],[],{"_220":2419,"_24":236,"_237":2420,"_239":2421},"6ea22a083826",[],"As he sought to ensure the project yielded reliable graph representations, Avagyan found that asking LLMs to one-shot graph representations didn’t quite work. “[The LLMs] hallucinated some relationship edges and nodes. So, the project uses Tree-sitter parsers for different languages, which give you the basic relationships you can extract from AST as something of a unified way of using any type of language.”","6154d54ed4e6",[2425],[],{"_220":2426,"_24":236,"_237":2427,"_239":2428},"1641bec46e7a",[],"Making Code Graphs Deterministic with AST","a36ee5a77fd2",[2432],[],{"_220":2433,"_24":236,"_237":2434,"_239":2435},"b5e47a104baa",[],"“Obviously, this has lots of benefits. If you ask a question about a specific function or a service with semantic search, it looks at the vector similarity and brings you some functions, some of which may have relevant names, but may not be relevant to your actual query. But using graphs, if you can translate your intent into Cypher (with the help of AI), you can basically pinpoint which function you are looking for.” The creator clarifies that the project also supports semantic search as an add-on to support things like code comments, and that in practice, graph parsing and semantic search tend to be complementary.","4b92136c2a79",[2439],[],{"_220":2440,"_24":236,"_237":2441,"_239":2442},"916fbfa96bd2",[],"Avagyan’s approach was to unify the variety within a single graph ontology that worked across multiple languages, from which he designed the project to let users map relationships within a monorepo, even those that contain multiple languages (the project currently supports more than a dozen). With such precise relationships properly mapped, users can use Cypher or another query language (or some sort of tool to map natural language to a query language) to query any part of a codebase as a graph, and compare against semantic search.","755c01bddb6f",[2452,2453,2454,2455,2456],[2446,2447],{"_220":2450,"_24":1959,"_1960":2451},{"_220":2448,"_24":1959,"_1960":2449},"249727404c78","https://en.wikipedia.org/wiki/Abstract_syntax_tree","d7c41e85c164","https://tree-sitter.github.io/tree-sitter/",{"_220":2469,"_24":236,"_237":2470,"_239":2471},{"_220":2466,"_24":236,"_237":2467,"_239":2468},{"_220":2463,"_24":236,"_237":2464,"_239":2465},{"_220":2460,"_24":236,"_237":2461,"_239":2462},{"_220":2457,"_24":236,"_237":2458,"_239":2459},"916d54525867",[]," and gives you all the syntactic relationships, it's not still enough because there’s so much variety between languages and frameworks.”","ecb02d6d29a7",[2448],"abstract syntax tree","2d414b50c68b",[],", which parses the ","c4f79c67c17a",[2450],"Tree-sitter","8cfad1aaed2e",[],"The creator clarifies that precisely mapping codebases to graphs requires a correct build for each framework and programming language. “Even with an underlying technology like ","cc2e0c67c8ad",[2475],[],{"_220":2476,"_24":236,"_237":2477,"_239":2478},"a7a96359e940",[],"Accounting for Variation Across Languages, Frameworks","55289533b05a",[2482],[],{"_220":2483,"_24":236,"_237":2484,"_239":2485},"494b6d9eb54a",[],"The creator suggests that the project specifically addresses the dichotomy between graph search and traditional text search/grep. “The first part of the project is how you go from code to graph. The second is an assistant that has a context of a graph, but it also has all the tools that typical coding agents or harnesses have. I also use it alongside my other coding harnesses. For some problems, it can find solutions faster and more accurately.”","8edb1e48b5a2",[2489],[],{"_220":2490,"_24":236,"_237":2491,"_239":2492},"b5cb7e6813d4",[],"The creator clarifies, “that doesn't mean you should use [text] or [graphs] exclusively. They can be used in conjunction with each other. You can have grep, other types of search, and graph search together. Then, given that the AI has both contexts, it will decide which way to go.”","6bfeff21d7b3",[2499,2500,2501],[2496],{"_220":2497,"_24":1959,"_1960":2498},"40d5a9d1fab0","https://en.wikipedia.org/wiki/Cypher_(query_language)",{"_220":2508,"_24":236,"_237":2509,"_239":2510},{"_220":2505,"_24":236,"_237":2506,"_239":2507},{"_220":2502,"_24":236,"_237":2503,"_239":2504},"850209620fae",[]," query that will calculate it for you in milliseconds.” The creator suggests that treating codebases as graphs in this way can conserve both tokens and time.","fd7434afa3fb",[2497],"Cypher","2814a2a422f6",[],"The creator gives the example of asking an agent to find the longest nested function call in your codebase. Today’s agents will search through the imports and invocations and do some counting, a task that LLMs aren’t great at. “But if you already have a graph that represents all core relationships between methods and functions, you or an AI agent can just write a ","c357da4c0347",[2517,2518,2519],[2514],{"_220":2515,"_24":1959,"_1960":2516},"67f4fbd2fe9a","https://en.wikipedia.org/wiki/Grep",{"_220":2526,"_24":236,"_237":2527,"_239":2528},{"_220":2523,"_24":236,"_237":2524,"_239":2525},{"_220":2520,"_24":236,"_237":2521,"_239":2522},"7de49dd818b7",[]," or possibly some kind of agentic search. As you might imagine, that can become inefficient, especially considering some of the questions you might typically have about a codebase.”","683e39e682aa",[2515],"grep","b42b4a2c8abb",[],"Avagyan notes that when you hand a codebase to one of today’s coding assistants, it’ll often treat it as a text artifact. “Then, it traverses the code base with a ","2fc0e81b9ef7",[2532],[],{"_220":2533,"_24":236,"_237":2534,"_239":2535},"4cb4e6406d24",[],"Viewing Code as a Retrievable Graph","b80310699127",[2542,2543],[2539],{"_220":2540,"_24":1959,"_1960":2541},"5b72c578e5d1","https://www.youtube.com/watch?v=O2jUTq6nCEY",{"_220":2546,"_24":236,"_237":2547,"_239":2548},{"_220":2544,"_24":236,"_237":2545,"_239":2335},"5ef047d4ccf9",[2540,2077],"a0931bc06a15",[2077],"Vitali Avagyan discusses using GraphRAG to parse repos with Memgraph. Image courtesy ","c9d37b6922a5",{"_1944":2551,"_24":1946},"image-bc24d7b1e62740bb7448a1a80a9dfd59940c68d6-3840x1100-jpg","9b432548c002",[2561,2562,2563,2564,2565],[2555,2556],{"_220":2559,"_24":1959,"_1960":2560},{"_220":2557,"_24":1959,"_1960":2558},"b6ce27c00ce8","https://github.com/vitali87/code-graph-rag","a50030cc5b04","https://www.linkedin.com/in/vitali-avagyan-a1566234/",{"_220":2578,"_24":236,"_237":2579,"_239":2580},{"_220":2575,"_24":236,"_237":2576,"_239":2577},{"_220":2572,"_24":236,"_237":2573,"_239":2574},{"_220":2569,"_24":236,"_237":2570,"_239":2571},{"_220":2566,"_24":236,"_237":2567,"_239":2568},"c24ce7265e1d",[],".","c66f9ef7a776",[2557],"code-graph-rag","a0f94686e2c4",[]," asserts that the key to understanding code, particularly code across different languages, is to stop thinking of it as lines of plain text, and to treat it as a relational graph, which is the basis of his open-source project ","7aa3b54f64f6",[2559],"Vitali Avagyan","9c50f1246f14",[],"As generating code becomes free, understanding what any of it does becomes much more costly, and updating it becomes significantly more complex. Developer ","4b5755f9b8bc",[2590,2591,2592,2593,2594],[2584,2585],{"_220":2588,"_24":1959,"_1960":2589},{"_220":2586,"_24":1959,"_1960":2587},"8042bdcd04fd","https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/cloud-next-2026-sundar-pichai/","afe823ab102f","https://techcrunch.com/2025/04/29/microsoft-ceo-says-up-to-30-of-the-companys-code-was-written-by-ai/",{"_220":2607,"_24":236,"_237":2608,"_239":2609},{"_220":2604,"_24":236,"_237":2605,"_239":2606},{"_220":2601,"_24":236,"_237":2602,"_239":2603},{"_220":2598,"_24":236,"_237":2599,"_239":2600},{"_220":2595,"_24":236,"_237":2596,"_239":2597},"70dd8d1c5fc1",[]," of their new code is AI-generated, which does nothing to help the ongoing challenges of first-time onboarding for codebases and the updating of existing codebases for newly-hired engineers and growing teams (as well as for agentic coding assistants).","a7022507758b",[2586],"75%","47cf0e8a331c",[]," to ","d1aebf2378c0",[2588],"30%","adc4d52ad4cd",[],"Major engineering orgs claim that as much as ","f09e9612ef15",[2613],[],{"_220":2614,"_24":236,"_237":2615,"_239":2616},"85de00f41251",[],"Why Code-as-Text Doesn’t Work for Genuine Understanding",[2623,2624,2625,2626,2627,2628,2629,2630,2631,2632,2633,2634,2635,2636,2637,2638,2639,2640,2641,2642,2643,2644,2645,2646,2647,2648,2649,2650,2651,2652,2653,2654,2655,2656,2657,2658,2659,2660,2661,2662,2663,2664,2665,2666,2667,2668,2669,2670,2671,2672,2673,2674,2675,2676,2677,2678,2679,2680,2681,2682,2683,2684,2685,2686,2687,2688,2689,2690,2691,2692,2693,2694,2695,2696,2697,2698,2699,2700,2701,2702,2703,2704,2705,2706,2707,2708,2709,2710],20,"7f2db4aa-8e93-45db-a467-a326acd92cfb","2026-10-02T19:09:00.000Z","long-horizon-agents-part-2-memory-is-more-than-recall","Long-Horizon Agents, Part 2: Memory Is More than Recall",{"_220":3830,"_24":222,"_223":3831,"_225":3832,"_227":576},{"_220":3824,"_24":222,"_223":3825,"_225":3826,"_227":228},{"_220":3806,"_24":222,"_223":3807,"_225":3808,"_227":228},{"_220":3799,"_24":222,"_223":3800,"_225":3801,"_227":228},{"_220":3792,"_24":222,"_223":3793,"_2727":2728,"_2729":2730,"_225":3794,"_227":228},{"_220":3785,"_24":222,"_223":3786,"_2727":2728,"_2729":2730,"_225":3787,"_227":228},{"_220":3778,"_24":222,"_223":3779,"_2727":2728,"_2729":2730,"_225":3780,"_227":228},{"_220":3763,"_24":222,"_223":3764,"_225":3765,"_227":228},{"_220":3756,"_24":222,"_223":3757,"_225":3758,"_227":576},{"_220":3749,"_24":222,"_223":3750,"_225":3751,"_227":228},{"_220":3742,"_24":222,"_223":3743,"_225":3744,"_227":228},{"_220":3735,"_24":222,"_223":3736,"_225":3737,"_227":228},{"_220":3728,"_24":222,"_223":3729,"_225":3730,"_227":244},{"_220":3722,"_24":222,"_223":3723,"_225":3724,"_227":244},{"_220":3715,"_24":222,"_223":3716,"_225":3717,"_227":244},{"_220":3708,"_24":222,"_223":3709,"_225":3710,"_227":244},{"_220":3701,"_24":222,"_223":3702,"_225":3703,"_227":244},{"_220":3694,"_24":222,"_223":3695,"_225":3696,"_227":244},{"_220":3687,"_24":222,"_223":3688,"_225":3689,"_227":244},{"_220":3681,"_24":222,"_223":3682,"_225":3683,"_227":244},{"_220":3674,"_24":222,"_223":3675,"_225":3676,"_227":244},{"_220":3667,"_24":222,"_223":3668,"_225":3669,"_227":244},{"_220":3660,"_24":222,"_223":3661,"_225":3662,"_227":244},{"_220":3653,"_24":222,"_223":3654,"_225":3655,"_227":244},{"_220":3646,"_24":222,"_223":3647,"_225":3648,"_227":244},{"_220":3639,"_24":222,"_223":3640,"_225":3641,"_227":244},{"_220":3632,"_24":222,"_223":3633,"_225":3634,"_227":244},{"_220":3625,"_24":222,"_223":3626,"_225":3627,"_227":228},{"_220":3618,"_24":222,"_223":3619,"_225":3620,"_227":228},{"_220":3601,"_24":222,"_223":3602,"_225":3603,"_227":228},{"_220":3598,"_24":1973,"_74":3599},{"_220":3580,"_24":222,"_223":3581,"_225":3582,"_227":228},{"_220":3573,"_24":222,"_223":3574,"_225":3575,"_227":228},{"_220":3566,"_24":222,"_223":3567,"_225":3568,"_227":576},{"_220":3559,"_24":222,"_223":3560,"_225":3561,"_227":228},{"_220":3537,"_24":222,"_223":3538,"_225":3539,"_227":228},{"_220":3507,"_24":222,"_223":3508,"_225":3509,"_227":228},{"_220":3448,"_24":222,"_223":3449,"_225":3450,"_227":228},{"_220":3430,"_24":222,"_223":3431,"_225":3432,"_227":228},{"_220":3423,"_24":222,"_223":3424,"_225":3425,"_227":576},{"_220":3416,"_24":222,"_223":3417,"_225":3418,"_227":228},{"_220":3367,"_24":222,"_223":3368,"_225":3369,"_227":228},{"_220":3360,"_24":222,"_223":3361,"_225":3362,"_227":228},{"_220":3333,"_24":222,"_223":3334,"_225":3335,"_227":228},{"_220":3311,"_24":222,"_223":3312,"_225":3313,"_227":228},{"_220":3284,"_24":222,"_223":3285,"_225":3286,"_227":228},{"_220":3259,"_24":222,"_223":3260,"_225":3261,"_227":228},{"_220":3252,"_24":222,"_223":3253,"_225":3254,"_227":576},{"_220":3245,"_24":222,"_223":3246,"_225":3247,"_227":228},{"_220":3234,"_24":222,"_223":3235,"_225":3236,"_227":228},{"_220":3202,"_24":222,"_223":3203,"_225":3204,"_227":228},{"_220":3195,"_24":222,"_223":3196,"_225":3197,"_227":228},{"_220":3188,"_24":222,"_223":3189,"_225":3190,"_227":244},{"_220":3181,"_24":222,"_223":3182,"_225":3183,"_227":244},{"_220":3174,"_24":222,"_223":3175,"_225":3176,"_227":244},{"_220":3167,"_24":222,"_223":3168,"_225":3169,"_227":244},{"_220":3160,"_24":222,"_223":3161,"_225":3162,"_227":228},{"_220":3149,"_24":222,"_223":3150,"_225":3151,"_227":228},{"_220":3138,"_24":222,"_223":3139,"_225":3140,"_227":228},{"_220":3093,"_24":222,"_223":3094,"_225":3095,"_227":228},{"_220":3086,"_24":222,"_223":3087,"_225":3088,"_227":576},{"_220":3079,"_24":222,"_223":3080,"_225":3081,"_227":228},{"_220":3072,"_24":222,"_223":3073,"_225":3074,"_227":2799},{"_220":3051,"_24":222,"_223":3052,"_2727":2728,"_2729":2730,"_225":3053,"_227":228},{"_220":3041,"_24":222,"_223":3042,"_2727":2728,"_2729":2730,"_225":3043,"_227":228},{"_220":3019,"_24":222,"_223":3020,"_2727":2728,"_2729":2730,"_225":3021,"_227":228},{"_220":3012,"_24":222,"_223":3013,"_225":3014,"_227":228},{"_220":3005,"_24":222,"_223":3006,"_225":3007,"_227":2799},{"_220":2984,"_24":222,"_223":2985,"_2727":2728,"_2729":2730,"_225":2986,"_227":228},{"_220":2974,"_24":222,"_223":2975,"_2727":2728,"_2729":2730,"_225":2976,"_227":228},{"_220":2964,"_24":222,"_223":2965,"_2727":2728,"_2729":2730,"_225":2966,"_227":228},{"_220":2946,"_24":222,"_223":2947,"_225":2948,"_227":228},{"_220":2939,"_24":222,"_223":2940,"_225":2941,"_227":2799},{"_220":2919,"_24":222,"_223":2920,"_2727":2728,"_2729":2730,"_225":2921,"_227":228},{"_220":2898,"_24":222,"_223":2899,"_2727":2728,"_2729":2730,"_225":2900,"_227":228},{"_220":2888,"_24":222,"_223":2889,"_2727":2728,"_2729":2730,"_225":2890,"_227":228},{"_220":2881,"_24":222,"_223":2882,"_225":2883,"_227":228},{"_220":2874,"_24":222,"_223":2875,"_225":2876,"_227":2799},{"_220":2853,"_24":222,"_223":2854,"_2727":2728,"_2729":2730,"_225":2855,"_227":228},{"_220":2832,"_24":222,"_223":2833,"_2727":2728,"_2729":2730,"_225":2834,"_227":228},{"_220":2811,"_24":222,"_223":2812,"_2727":2728,"_2729":2730,"_225":2813,"_227":228},{"_220":2804,"_24":222,"_223":2805,"_225":2806,"_227":228},{"_220":2796,"_24":222,"_223":2797,"_225":2798,"_227":2799},{"_220":2774,"_24":222,"_223":2775,"_2727":2728,"_2729":2730,"_225":2776,"_227":228},{"_220":2752,"_24":222,"_223":2753,"_2727":2728,"_2729":2730,"_225":2754,"_227":228},{"_220":2725,"_24":222,"_223":2726,"_2727":2728,"_2729":2730,"_225":2731,"_227":228},{"_220":2718,"_24":222,"_223":2719,"_225":2720,"_227":228},{"_220":2711,"_24":222,"_223":2712,"_225":2713,"_227":228},"b1115aef2063",[2714],[],{"_220":2715,"_24":236,"_237":2716,"_239":2717},"cc39ba41ca8f",[],"\n","d5296f448a9f",[2721],[],{"_220":2722,"_24":236,"_237":2723,"_239":2724},"8d1da74a6c00",[],"Mubit’s system is able to prioritize procedural guidance over raw history using typing and outcome signals. This is a good benefit that depends on reliable attribution, since even a successful deployment doesn’t necessarily prove that every retrieved lesson contributed to it.","828f0b6c00db",[2735,2736,2737,2738],"level",1,"listItem","bullet",[2732],{"_220":2733,"_24":1959,"_1960":2734},"b7f12682eee8","https://docs.mubit.ai/concepts/memory-model",{"_220":2748,"_24":236,"_237":2749,"_239":2750},{"_220":2745,"_24":236,"_237":2746,"_239":2747},{"_220":2742,"_24":236,"_237":2743,"_239":2744},{"_220":2739,"_24":236,"_237":2740,"_239":2741},"cf1e8502645e",[]," includes facts, traces, lessons, rules, workflows, and exact archives, with a separate, mutable working state (rather than specifying a particular database backend).","a124afaedcbc",[2733],"memory model","c220cbdd3ded",[]," Mubit’s documented ","a1974bc579f7",[2751],"Store:","strong","1ed494cdb84c",[2758,2759,2760,2761],[2755],{"_220":2756,"_24":1959,"_1960":2757},"03f6430bd2bb","https://docs.mubit.ai/concepts/retrieval-semantics",{"_220":2771,"_24":236,"_237":2772,"_239":2773},{"_220":2768,"_24":236,"_237":2769,"_239":2770},{"_220":2765,"_24":236,"_237":2766,"_239":2767},{"_220":2762,"_24":236,"_237":2763,"_239":2764},"82dc20558a47",[]," to combines semantic, lexical and temporal signals with working state and lesson overlays. The platform ranks information based on outcome history, and also has a direct mode that returns evidence without model-based routing or answer synthesis.","b175454eee20",[2756],"recall queries","4e86e92fe9b1",[]," The platform uses ","4390006d4807",[2751],"Read:","7f8a916d5d27",[2780,2781,2782,2783],[2777],{"_220":2778,"_24":1959,"_1960":2779},"1db8c126f415","https://docs.mubit.ai/concepts/learning-loop",{"_220":2793,"_24":236,"_237":2794,"_239":2795},{"_220":2790,"_24":236,"_237":2791,"_239":2792},{"_220":2787,"_24":236,"_237":2788,"_239":2789},{"_220":2784,"_24":236,"_237":2785,"_239":2786},"ab2045bc4caf",[]," that attaches outcomes to memory identifiers and uses reflection to derive conditional lessons for validation and promotion.","60a47075eb73",[2778],"learning loop","36a26a2ad674",[]," Mubit uses asynchronous ingestion to classify typed entries. The platform has a ","4a28c2e300ff",[2751],"Write:","23586738a3bd",[2800],[],"h3",{"_220":2801,"_24":236,"_237":2802,"_239":2803},"55a06b57abcf",[],"Mubit","66730d013fe4",[2807],[],{"_220":2808,"_24":236,"_237":2809,"_239":2810},"380b861a865e",[],"cognee’s approach lets teams customize the process of transforming source data to searchable memory. However, the approach also uses multiple derived representations, which creates extra maintenance overhead. Specifically, failed jobs can leave indexes incomplete which means not every dependent record instantly auto-updates with any corrections.","f777badb59d5",[2817,2818,2819,2820],[2814],{"_220":2815,"_24":1959,"_1960":2816},"0b959dcc7d95","https://docs.cognee.ai/core-concepts/architecture",{"_220":2830,"_24":236,"_237":2831,"_239":2750},{"_220":2827,"_24":236,"_237":2828,"_239":2829},{"_220":2824,"_24":236,"_237":2825,"_239":2826},{"_220":2821,"_24":236,"_237":2822,"_239":2823},"98a98865c12c",[]," combines relational records (for source metadata and provenance), vectors (for semantic search), and a graph (for entities and relationships).","8006ac452a9b",[2815],"architecture","b1af73cfdbec",[]," cognee’s ","3a786e68b48e",[2751],"0c8ba15851e3",[2838,2839,2840,2841],[2835],{"_220":2836,"_24":1959,"_1960":2837},"0d0d58abefcd","https://docs.cognee.ai/core-concepts/main-operations/recall",{"_220":2851,"_24":236,"_237":2852,"_239":2773},{"_220":2848,"_24":236,"_237":2849,"_239":2850},{"_220":2845,"_24":236,"_237":2846,"_239":2847},{"_220":2842,"_24":236,"_237":2843,"_239":2844},"e398196f7eb2",[]," supports different modes according to the evidence/answer required.","a7f71645af37",[2836],"recall interface","b28cafe0f3f1",[]," cognee’s retrieval uses vectors and graph structure. The ","f045c19eb548",[2751],"dcfbde86d09f",[2859,2860,2861,2862],[2856],{"_220":2857,"_24":1959,"_1960":2858},"4171b8028864","https://docs.cognee.ai/core-concepts/main-operations/improve",{"_220":2872,"_24":236,"_237":2873,"_239":2795},{"_220":2869,"_24":236,"_237":2870,"_239":2871},{"_220":2866,"_24":236,"_237":2867,"_239":2868},{"_220":2863,"_24":236,"_237":2864,"_239":2865},"5331b3438bf8",[]," that incorporate useful session information into the permanent graph.","dab6e9a6f8ee",[2857],"improvement passes","0e9444b094e5",[]," cognee uses configurable pipelines to ingest data, extract entities and relationships, and enrich existing memory. Later, the platform can also run ","9a40e96ce730",[2751],"844cbdc9c89b",[2877],[],{"_220":2878,"_24":236,"_237":2879,"_239":2880},"004318b7238a",[],"cognee","a9dca42167a7",[2884],[],{"_220":2885,"_24":236,"_237":2886,"_239":2887},"abf37fd05bec",[],"Graphiti’s approach to resolving changes during ingestion makes applicable state more explicit at read time. The trade-off here is that correct extraction becomes much more important. If the agent merges distinct workers or invalidates the wrong relationship, it may also corrupt subsequent retrieval.","c89ee678066d",[2891,2892],[],{"_220":2896,"_24":236,"_237":2897,"_239":2750},{"_220":2893,"_24":236,"_237":2894,"_239":2895},"1b9a86a13f83",[]," Graphiti uses a variety of artifacts for storage. Episodes retain source evidence, graph nodes represent entities, and edges represent relationships with temporal validity. (It should be noted that Graphiti is the open-source framework and Zep provides managed context-graph infrastructure.)","b05b70c9aa79",[2751],"3417dd3e4105",[2904,2905,2906,2907],[2901],{"_220":2902,"_24":1959,"_1960":2903},"bb5a4597ff40","https://help.getzep.com/graphiti/working-with-data/searching",{"_220":2917,"_24":236,"_237":2918,"_239":2773},{"_220":2914,"_24":236,"_237":2915,"_239":2916},{"_220":2911,"_24":236,"_237":2912,"_239":2913},{"_220":2908,"_24":236,"_237":2909,"_239":2910},"a5b1d580b210",[]," approach that combines semantic and BM25 retrieval with rank fusion. It also offers configurable recipes, which add graph traversal and node-distance ranking/cross-encoder reranking.","c45b531db105",[2902],"hybrid search","28d2a41d4e5d",[]," The platform uses a ","788181f131ab",[2751],"200e943d74d2",[2925,2926,2927,2928],[2922],{"_220":2923,"_24":1959,"_1960":2924},"44a39f9fb87c","https://github.com/getzep/graphiti",{"_220":2937,"_24":236,"_237":2938,"_239":2795},{"_220":2935,"_24":236,"_237":2936,"_239":279},{"_220":2932,"_24":236,"_237":2933,"_239":2934},{"_220":2929,"_24":236,"_237":2930,"_239":2931},"bd37a5647c0f",[]," extracts entities and relationships from incoming episodes. The platform can preserve evidence history but allows new evidence to invalidate previous relationships.","646ac358d278",[2923],"Graphiti","2bf68a99ad63",[],"23ab9ad7f341",[2751],"ec1a6ae1873c",[2942],[],{"_220":2943,"_24":236,"_237":2944,"_239":2945},"4d803c10905d",[],"Graphiti and Zep","27fdcaf80cf2",[2952,2953,2954],[2949],{"_220":2950,"_24":1959,"_1960":2951},"39e749b78ea5","https://github.com/mem0ai/mem0",{"_220":2961,"_24":236,"_237":2962,"_239":2963},{"_220":2958,"_24":236,"_237":2959,"_239":2960},{"_220":2955,"_24":236,"_237":2956,"_239":2957},"de1e3d5a2c9a",[]," that differs from the managed architecture described above.)","e5f17f165be9",[2950],"open-source implementation","95b36f2da3c8",[],"Mem0’s approach uses a standardized pipeline to interpret information during an agentic run. The platform takes the approach of preserving evidence by appending changed facts, but it leaves version resolution up to retrieval and reader. (Mem0 also publishes an ","98d3de1d13c4",[2967,2968],[],{"_220":2972,"_24":236,"_237":2973,"_239":2750},{"_220":2969,"_24":236,"_237":2970,"_239":2971},"3cc6a355c602",[]," Memory text, embeddings, and metadata live in a vector database. Mem0 uses an entity store to link related memories, and retains additional history and a rolling message window using SQL.","0dfb36a8c102",[2751],"39c6e15e2909",[2977,2978],[],{"_220":2982,"_24":236,"_237":2983,"_239":2773},{"_220":2979,"_24":236,"_237":2980,"_239":2981},"1711f15ec7f4",[]," Mem0 ranks information relevance based on semantic, BM25, entity, and temporal signals. Temporal relevance ranks information by freshness rather than immediately excluding stale candidates.","99f313c464e3",[2751],"6625977f2d5a",[2990,2991,2992,2993],[2987],{"_220":2988,"_24":1959,"_1960":2989},"d6a326566adf","https://docs.mem0.ai/core-concepts/memory-evaluation",{"_220":3003,"_24":236,"_237":3004,"_239":2795},{"_220":3000,"_24":236,"_237":3001,"_239":3002},{"_220":2997,"_24":236,"_237":2998,"_239":2999},{"_220":2994,"_24":236,"_237":2995,"_239":2996},"6d3effc41325",[]," processes conversations asynchronously, extracts ADD-only facts, deduplicates them, creates embeddings, and adds entity links and temporal metadata.","60b128b85711",[2988],"extraction pipeline","724c2c891981",[]," Mem0’s ","5a484c730a59",[2751],"813631713196",[3008],[],{"_220":3009,"_24":236,"_237":3010,"_239":3011},"d9f878ff27a0",[],"Mem0's Managed Platform","031f13e60a30",[3015],[],{"_220":3016,"_24":236,"_237":3017,"_239":3018},"7b3d106bb226",[],"Letta’s approach gives capable models flexibility over both representation and retrieval. Because Letta allows for persistent memory editing and on-demand retrieval, the correctness of a job depends on editing and search decisions. Version control provides rollback capability, but does not validate the meaning of a memory update.","0aaa75402c19",[3025,3026,3027,3028],[3022],{"_220":3023,"_24":1959,"_1960":3024},"bab345521546","https://www.letta.com/blog/memory-blocks/",{"_220":3038,"_24":236,"_237":3039,"_239":3040},{"_220":3035,"_24":236,"_237":3036,"_239":3037},{"_220":3032,"_24":236,"_237":3033,"_239":3034},{"_220":3029,"_24":236,"_237":3030,"_239":3031},"28d8492c3fe9",[],", which persist labeled, size-limited text. Context repositories represent memory as versioned files, with designated files loaded into the prompt.","2e258a11f020",[3023,2077],"memory blocks","bd33ca7d5b2b",[],"Letta abstracts context window storage as ","3a82d6616947",[2751],"Store: ","f63a7efb1736",[3044,3045],[],{"_220":3049,"_24":236,"_237":3050,"_239":2773},{"_220":3046,"_24":236,"_237":3047,"_239":3048},"c6ee02242da8",[]," Letta uses a combination of automatically including records and using agents to access additional memory through tools. In context repositories, the file hierarchy helps it discover material to load on demand.","225fb3f5614e",[2751],"d1ca287a4e65",[3057,3058,3059,3060],[3054],{"_220":3055,"_24":1959,"_1960":3056},"d611832c5bfa","https://www.letta.com/blog/context-repositories/",{"_220":3070,"_24":236,"_237":3071,"_239":2795},{"_220":3067,"_24":236,"_237":3068,"_239":3069},{"_220":3064,"_24":236,"_237":3065,"_239":3066},{"_220":3061,"_24":236,"_237":3062,"_239":3063},"b30686ec4f7d",[]," let memory agents work in isolated Git worktrees and merge their changes.","2162f1f7e868",[3055],"context repositories","7f1857d4cccd",[]," Agents edit persistent memory using tools, with background agents handling consolidation. Letta Code's ","e107ebffb775",[2751],"0893565179df",[3075],[],{"_220":3076,"_24":236,"_237":3077,"_239":3078},"606ad4cb319c",[],"Letta and Letta Code","4f649d90264a",[3082],[],{"_220":3083,"_24":236,"_237":3084,"_239":3085},"245d14e460fa",[],"Below are some real-world examples of how different platforms approach agentic memory, at least, according to publicly-available source code or architecture explainers:","4747fa21e7a7",[3089],[],{"_220":3090,"_24":236,"_237":3091,"_239":3092},"a0bf90cc4808",[],"How Existing Systems Combine These Choices","26e03d89cef4",[3102,3103,3104,3105,3106,3107,3108,3109,3110],[3096,3097],{"_220":3100,"_24":1959,"_1960":3101},{"_220":3098,"_24":1959,"_1960":3099},"c09991de458c","https://arxiv.org/abs/2512.13564","580336df6328","https://www.ibm.com/docs/en/db2/12.1.x?topic=indexes-vector",{"_220":3135,"_24":236,"_237":3136,"_239":3137},{"_220":3132,"_24":236,"_237":3133,"_239":3134},{"_220":3129,"_24":236,"_237":3130,"_239":3131},{"_220":3126,"_24":236,"_237":3127,"_239":3128},{"_220":3123,"_24":236,"_237":3124,"_239":3125},{"_220":3120,"_24":236,"_237":3121,"_239":3122},{"_220":3117,"_24":236,"_237":3118,"_239":3119},{"_220":3114,"_24":236,"_237":3115,"_239":3116},{"_220":3111,"_24":236,"_237":3112,"_239":3113},"2fbdca59fcea",[],".)","c88f2f3993b8",[3098],"Memory in the Age of AI Agents","241c4055abee",[]," determine whether facts are current, whether records refer to the same worker, or whether a proposed lesson is justified. For example, selection, organization, and retrieval seem like a huge part of memory design for Claude Code, even when the durable records are ordinary files. (For more on latent and parametric memory, see the report ","592bc7c5d608",[2077],"not","f69b2c969e9a",[],"different representations to find similar content. However, vector indexes do ","51dc1c62d004",[2077],"alongside ","4f49328c1072",[]," (specialized data structures that organize embeddings to speed up searches) can sit ","bfff0f664909",[3100,2077],"Vector indexes","65b134718ba7",[2077],"Note: ","459ca3e10e7d",[3141,3142],[],{"_220":3146,"_24":236,"_237":3147,"_239":3148},{"_220":3143,"_24":236,"_237":3144,"_239":3145},"9580a2df3917",[]," Using graphs could connect customers to worker fleets, fleets to software versions, and versions to schema requirements. A key feature of graphs is temporal relationships, which could indicate when each dependency applied during the run. With graphs, multi-step questions (like which customers still depend on a particular column) become easier to express. However, as underlying deployments change, graphs require extraction, entity resolution, and updates to remain accurate.","6597cc79ccc4",[2751],"Entity-relationship graphs.","91041a2f7d82",[3152,3153],[],{"_220":3157,"_24":236,"_237":3158,"_239":3159},{"_220":3154,"_24":236,"_237":3155,"_239":3156},"048b0f97261a",[]," If you need to answer questions across many customers and deployments, another approach is using rows or documents for important fields like customer_id, worker_version, observed_at, and source_event_id. Using fields lets agents filter and make explicit version checks. However, you would need a schema that tracks version changes among workers, which is a distinction that can be lost in a field overwrite unless you also retain earlier versions.","c4107be75da4",[2751],"Structured records.","f32fca39f7ed",[3163],[],{"_220":3164,"_24":236,"_237":3165,"_239":3166},"cfea8479c38f",[],"The index helps discover the notes, which explain the constraint and point to its source. Using ordinary file tools makes such summaries easy to inspect and revise. Also, files can contain structured metadata, even if our system chooses to use Markdown. The challenge here is consistency, since notes can be readable while contradicting each other. Adding versioning makes changes reversible, but does not validate them. (See Letta Code, below, for an example.)","06c83b269941",[3170],[],{"_220":3171,"_24":236,"_237":3172,"_239":3173},"aa317516e315",[],"└── reference_customer_rollout_tracker.md","e6d70c556e93",[3177],[],{"_220":3178,"_24":236,"_237":3179,"_239":3180},"1b55f6b25b56",[],"├── feedback_migration_compatibility.md","9d066e09a773",[3184],[],{"_220":3185,"_24":236,"_237":3186,"_239":3187},"baf15381d1a5",[],"├── MEMORY.md","92b2b15e6ef5",[3191],[],{"_220":3192,"_24":236,"_237":3193,"_239":3194},"593cd17fb5cc",[],"memory/","989c53d303db",[3198],[],{"_220":3199,"_24":236,"_237":3200,"_239":3201},"9091690fe872",[],"Here is an illustrative layout for our example:","586a0cf2fe77",[3211,3212,3213,3214,3215,3216],[3205,3206],{"_220":3209,"_24":1959,"_1960":3210},{"_220":3207,"_24":1959,"_1960":3208},"bf2bc5254bf0","https://code.claude.com/docs/en/memory#storage-location","961db163e8ee","https://code.claude.com/docs/en/skills",{"_220":3231,"_24":236,"_237":3232,"_239":3233},{"_220":3228,"_24":236,"_237":3229,"_239":3230},{"_220":3225,"_24":236,"_237":3226,"_239":3227},{"_220":3222,"_24":236,"_237":3223,"_239":3224},{"_220":3219,"_24":236,"_237":3220,"_239":3221},{"_220":3217,"_24":236,"_237":3218,"_239":2568},"84bc558b7fb4",[],"9b5977a7be50",[3207],"default memory is local to the machine","8411ca9c5db5",[],", but the ","3dbe26891ce6",[3209,2077],"frontmatter","0f9e248beec2",[]," Claude Code stores auto memory as Markdown under ~/.claude/projects//memory/, with a MEMORY.md index and individual topic files. Different memory types appear in ","63ac9cbc656e",[2751],"Text files and summaries.","221d0943bbbe",[3237,3238],[],{"_220":3242,"_24":236,"_237":3243,"_239":3244},{"_220":3239,"_24":236,"_237":3240,"_239":3241},"7c4441ab063c",[]," Claude Code's session transcripts show that the model retains interactions. (Files or object storage can also hold command output, patches, etc..) Such artifacts help agents investigate and reprocess, but even a full history does not guarantee you can immediately identify which constraint applies to the current job. For that, you would need your agents to run a subsequent read or derived view.","8961e61692ba",[2751],"Raw logs and artifacts.","4ce4eb451d82",[3248],[],{"_220":3249,"_24":236,"_237":3250,"_239":3251},"07928571dd78",[],"There are multiple ways to represent an agent’s experience, including an execution record, a written lesson, a structured fact, or a relationship. Your choice determines what aspects of an agentic job you can preserve and query directly. (Your backend determines how records persist and how to coordinate concurrent access.) We’ll cover four approaches below:","1adfa149433a",[3255],[],{"_220":3256,"_24":236,"_237":3257,"_239":3258},"788c87082178",[],"Storage: Representation, Backend, and Indexes","7d7e734abb9d",[3268,3269,3270,3271],[3262,3263],{"_220":3266,"_24":1959,"_1960":3267},{"_220":3264,"_24":1959,"_1960":3265},"067a18a9a1e9","https://code.claude.com/docs/en/how-claude-code-works#when-context-fills-up","1af973a49559","https://learn.microsoft.com/en-us/agent-framework/concepts/agents/conversations/compaction",{"_220":3281,"_24":236,"_237":3282,"_239":3283},{"_220":3278,"_24":236,"_237":3279,"_239":3280},{"_220":3275,"_24":236,"_237":3276,"_239":3277},{"_220":3272,"_24":236,"_237":3273,"_239":3274},"856e85a5c3db",[],", supporting continuity within a session while treating the selection of durable lessons for future sessions as a separate decision.","66110f9314ae",[3264],"summarize a growing conversation within its context window","49dd1e323c26",[]," (condensing older conversation history, reasoning, and tool outputs to conserve context window) can also help here. Claude Code’s approach is to ","79c644312837",[3266,2077],"Compaction","1d284f0bdd3e",[3287,3288,3289,3290,3291,3292],[],{"_220":3308,"_24":236,"_237":3309,"_239":3310},{"_220":3305,"_24":236,"_237":3306,"_239":3307},{"_220":3302,"_24":236,"_237":3303,"_239":3304},{"_220":3299,"_24":236,"_237":3300,"_239":3301},{"_220":3296,"_24":236,"_237":3297,"_239":3298},{"_220":3293,"_24":236,"_237":3294,"_239":3295},"b5f67f56504d",[],". It’s possible for both automatic or agentic-requested read requests to use exact lookups, keyword search, vector similarity, or graph traversal. (Harnesses can also combine policies like including a small index while auto-retrieving known task constraints and letting agents investigate further.) The goal is to select enough evidence for decision-making without having to store the entire history into every call.","faf7027e81fd",[2077],"how retrieval works","17b0e32a878e",[]," initiates retrieval is separate from ","9c22117de48b",[2077],"Who","d277a5a24136",[],": ","8d32937852d7",[2077],"About retrieval and memory overhead","baec8a07be2b",[3317,3318,3319,3320],[3314],{"_220":3315,"_24":1959,"_1960":3316},"0f84a55baee9","https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents",{"_220":3330,"_24":236,"_237":3331,"_239":3332},{"_220":3327,"_24":236,"_237":3328,"_239":3329},{"_220":3324,"_24":236,"_237":3325,"_239":3326},{"_220":3321,"_24":236,"_237":3322,"_239":3323},"413c9af36bc8",[],", so the model can discover a clue, refine its question, and perform additional retrieval. Agents using this policy might start by reading a rollout note, follow its reference to the tracker, and check current worker versions before proposing a migration (or, if the task was about why the original migration failed, it would search retained session records for more details). This approach adds flexibility but also adds costs for model calls and tool use. Having a compact index can help an investigation start better, but isn’t a guarantee it will end well, as agents may fail to search, choose the wrong note, or stop short of checking the source.","bfef448b405c",[3315],"Claude Code combines upfront context with on-demand retrieval","de9d369d54d7",[]," Another approach is to let agents use tools to investigate. For example, ","5db73706b4b4",[2751],"Let the model retrieve memory as needed.","3a95128b53ea",[3336,3337,3338,3339,3340,3341],[],{"_220":3357,"_24":236,"_237":3358,"_239":3359},{"_220":3354,"_24":236,"_237":3355,"_239":3356},{"_220":3351,"_24":236,"_237":3352,"_239":3353},{"_220":3348,"_24":236,"_237":3349,"_239":3350},{"_220":3345,"_24":236,"_237":3346,"_239":3347},{"_220":3342,"_24":236,"_237":3343,"_239":3344},"4f39fb880af4",[],". However, this policy carries the risk that the harness might introduce new issues (like constructing a poor query, returning stale records, or filling context with superficially similar stuff).","1c9c00d144c4",[2077],"think to search","d71331d31241",[]," selected model calls, without waiting for the model to request it. This policy requires a migration-aware harness to proactively identify any affected services/schemas, retrieve any relevant deployment constraints, and attach them to the next call. (This approach would support exact filters with structured records and allow free-form notes to be text-searched.) The benefit here is that retrieval is not constrained by whether agents ","58a66df2bcde",[2077],"before","8406b79505b0",[]," Another approach runs retrieval ","7061a6e00d28",[2751],"Have the harness retrieve and assemble context.","da3b59765f18",[3363],[],{"_220":3364,"_24":236,"_237":3365,"_239":3366},"8a34dbf8511d",[],"Automatic inclusion makes discovery more predictable, but consumes context even on unrelated tasks. It also requires a choice about what deserves that space. CLAUDE.md supplies persistent instructions; auto memory contains Claude's own notes. Both guide behavior rather than enforce it. A deployment constraint that must be guaranteed needs an execution check as well as a reminder. Claude Code's documentation explains this distinction.","5ca9d96310af",[3376,3377,3378,3379,3380,3381,3382,3383,3384,3385],[3370,3371],{"_220":3374,"_24":1959,"_1960":3375},{"_220":3372,"_24":1959,"_1960":3373},"2fe7d1e0fb7c","https://code.claude.com/docs/en/memory#claudemd-vs-auto-memory","6b5816824a71","https://code.claude.com/docs/en/memory#how-it-works",{"_220":3413,"_24":236,"_237":3414,"_239":3415},{"_220":3410,"_24":236,"_237":3411,"_239":3412},{"_220":3407,"_24":236,"_237":3408,"_239":3409},{"_220":3404,"_24":236,"_237":3405,"_239":3406},{"_220":3401,"_24":236,"_237":3402,"_239":3403},{"_220":3398,"_24":236,"_237":3399,"_239":3400},{"_220":3395,"_24":236,"_237":3396,"_239":3397},{"_220":3392,"_24":236,"_237":3393,"_239":3394},{"_220":3389,"_24":236,"_237":3390,"_239":3391},{"_220":3386,"_24":236,"_237":3387,"_239":3388},"f04028db5006",[]," between persistent instructions vs. auto memory.)","72cea0fbced3",[3372],"Claude Code distinguishes","0ca414d6225d",[],"because a concise and descriptive entry makes the right notes easier to discover (and bad descriptions make it harder). This approach makes discovery more predictable but consumes context on unrelated tasks, and requires a choice about what is worth storing. (For reference, ","0301e86aad6f",[2077],"write ","a7474d19202d",[]," with ","d9ba8f226466",[2077],"read","53bf244436b4",[],"). Here, record storage does not count against total stored memory, as any reading of detailed topic files is done on demand. This approach connects ","858d0ef5328a",[3374],"loads MEMORY.md at conversation start","351d6bf3467b",[]," One approach is to keep a small set of instructions or pointers within context (Claude Code ","81afda08f68e",[2751],"Include designated records automatically.","0d2f5b0ff899",[3419],[],{"_220":3420,"_24":236,"_237":3421,"_239":3422},"b87351817f0c",[],"After noting changes from their write policy, agents need to understand whether written changes affect them and any upgrades they have received. These three read policies are most common:","330c4b5540cf",[3426],[],{"_220":3427,"_24":236,"_237":3428,"_239":3429},"978e8914fa68",[],"Read Policy: What Reaches the Model?","28f6819109be",[3433,3434,3435,3436],[],{"_220":3445,"_24":236,"_237":3446,"_239":3447},{"_220":3442,"_24":236,"_237":3443,"_239":3444},{"_220":3440,"_24":236,"_237":3441,"_239":3122},{"_220":3437,"_24":236,"_237":3438,"_239":3439},"17bde65913e4",[]," mutually exclusive. For example, you could retain an exchange, add a brief reference, and reconcile it with other experiences later. Of course, you would need to account for delays if subsequent processing happens in the background. Conversely, you would have to account for newly-discovered constraints if you ran immediate retries.","85f7c1ab0cff",[2077],"2635523039da",[]," These write policies are ","fcd21325e323",[2077],"Note:","5861fa445fa5",[3459,3460,3461,3462,3463,3464,3465,3466,3467,3468,3469,3470],[3451,3452,3453],{"_220":3457,"_24":1959,"_1960":3458},{"_220":3455,"_24":1959,"_1960":3456},{"_220":3454,"_24":1959,"_1960":3056},"cf05d9d85127","94e81cff0ae0","https://www.ibm.com/docs/en/informix-servers/12.10.0?topic=modeling-fact-table","b1a64c40a930","https://en.wikipedia.org/wiki/Data_extraction",{"_220":3504,"_24":236,"_237":3505,"_239":3506},{"_220":3501,"_24":236,"_237":3502,"_239":3503},{"_220":3498,"_24":236,"_237":3499,"_239":3500},{"_220":3495,"_24":236,"_237":3496,"_239":3497},{"_220":3492,"_24":236,"_237":3493,"_239":3494},{"_220":3489,"_24":236,"_237":3490,"_239":3491},{"_220":3486,"_24":236,"_237":3487,"_239":3488},{"_220":3483,"_24":236,"_237":3484,"_239":3485},{"_220":3480,"_24":236,"_237":3481,"_239":3482},{"_220":3477,"_24":236,"_237":3478,"_239":3479},{"_220":3474,"_24":236,"_237":3475,"_239":3476},{"_220":3471,"_24":236,"_237":3472,"_239":3473},"91234870a812",[]," for documented examples of using background consolidation to update agent memory.","938fcb5e36a8",[3454],"Letta Code's context repositories","b37a5929e4ab",[]," (deciding whether new facts should update an existing memory). See ","5e3b514ea686",[2077],"consolidation","cdd413341cbd",[],") from ","f07d4da484cd",[3455],"candidate facts","00c26c30913f",[]," (used to identify ","ce486f61cec6",[2077],"extraction","8f6092b664db",[],". Converting the resulting data from the job into appropriate data fields provides a consistent schema and lets us rerun extraction when the policy changes. However, the extra processing adds cost and potential risk of misinterpreting evidence. It also separates ","efeadcbf2b1c",[3457,2077],"extractor","c729269fd4b9",[]," Another approach is processing recorded sessions independently of the agent with an ","097ba3fb5001",[2751],"Use a separate extraction pipeline.","b0388f909a33",[3513,3514,3515,3516,3517,3518],[3510],{"_220":3511,"_24":1959,"_1960":3512},"86369451d645","https://code.claude.com/docs/en/memory#auto-memory",{"_220":3534,"_24":236,"_237":3535,"_239":3536},{"_220":3531,"_24":236,"_237":3532,"_239":3533},{"_220":3528,"_24":236,"_237":3529,"_239":3530},{"_220":3525,"_24":236,"_237":3526,"_239":3527},{"_220":3522,"_24":236,"_237":3523,"_239":3524},{"_220":3519,"_24":236,"_237":3520,"_239":3521},"ed7d89ec8a3e",[],"saving (such as which data column changed), while the operating constraint, and where to verify it, becomes durable. This approach requires less reconstruction on the next run. However, there’s a risk of misinterpreting a one-off issue, like a customer delay, as a hard-and-fast rule to implement (which can be mitigated with notes/references to supporting evidence).","58fd06e03e07",[2077],"worth ","a377237b09e8",[]," selectively saves user info, feedback, context, and references, skipping facts recoverable from code or Git history. The difference is what is ","ea35d09d64e9",[3511],"auto memory","d0750afe804b",[]," Claude Code's ","a3d9b929d5ec",[2751],"Let the active agent write a memory.","d809f9f408ed",[3543,3544,3545,3546],[3540],{"_220":3541,"_24":1959,"_1960":3542},"f5a4020c4fb9","https://code.claude.com/docs/en/how-claude-code-works#work-with-sessions",{"_220":3556,"_24":236,"_237":3557,"_239":3558},{"_220":3553,"_24":236,"_237":3554,"_239":3555},{"_220":3550,"_24":236,"_237":3551,"_239":3552},{"_220":3547,"_24":236,"_237":3548,"_239":3549},"c0a92cd924e3",[],". Recording means preserving a failed query + investigation + any dev explanation, but does not require immediate decisions on which details matter. Auto-capture moves the cost of the write policy to read, as someone will need to search for and interpret it later (which would mean that this approach also requires retention).","de3b99a10fc3",[3541],"records messages, tool calls, and results","3c08d4da140a",[]," Claude Code ","d53459c4b154",[2751],"Automatically capture execution.","ea5d8e984d9d",[3562],[],{"_220":3563,"_24":236,"_237":3564,"_239":3565},"6cedcb2c785b",[],"A write policy determines what gets retained, who interprets it, and when the result becomes available. These three are most common:","ccf4da4770a3",[3569],[],{"_220":3570,"_24":236,"_237":3571,"_239":3572},"424c91921778",[],"Write Policy: What Becomes Memory?","d676ce5bde85",[3576],[],{"_220":3577,"_24":236,"_237":3578,"_239":3579},"94f5b2979df1",[],"To follow these choices through one task, imagine using Claude Code to investigate a database migration that broke an older worker. During the investigation, the developer explains that customers upgrade their workers independently of the backend, and that a rollout tracker records which versions are still running. What should survive it, and how should that information affect the next migration?","26defa59f442",[3586,3587,3588],[3583],{"_220":3584,"_24":1959,"_1960":3585},"e04bb9c15fd2","https://x.com/himanshustwts/status/2038924027411222533",{"_220":3595,"_24":236,"_237":3596,"_239":3597},{"_220":3592,"_24":236,"_237":3593,"_239":3594},{"_220":3589,"_24":236,"_237":3590,"_239":3591},"921f830b5e17",[2077],". This is a source-code analysis, not an official architecture diagram. Internal mechanisms such as autoDream and extractMemories reflect the version analyzed. The walkthrough below uses public documentation for shipped behavior.","9767f044ab49",[3584,2077],"himanshu on X, March 31, 2026","083a41cf889e",[2077],"Source: ","ad95b655bc4d",{"_1944":3600,"_24":1946},"image-f8350e46e885d0b2fa3ac0a247b0dbef59722d0b-1200x720-jpg","6196452f03db",[3606,3607,3608],[3604],{"_220":3605,"_24":1959,"_1960":3585},"b6eb1cd451c2",{"_220":3615,"_24":236,"_237":3616,"_239":3617},{"_220":3612,"_24":236,"_237":3613,"_239":3614},{"_220":3609,"_24":236,"_237":3610,"_239":3611},"4b7379850874",[]," while reading about memory systems. The diagram separates the paths that create and revise memory from the paths that bring it back into context:","dc5e6e608e02",[3605],"himanshu's walkthrough of Claude Code's memory architecture","3a123eb6cc1c",[],"I came across ","7abfa2abb045",[3621],[],{"_220":3622,"_24":236,"_237":3623,"_239":3624},"09e9d2285318",[],"The implementation of memory comes down to three connected choices: the write policy, the read policy, and the storage representation. Claude Code gives us a concrete example of how a harness combines them.","a70d3dc5602d",[3628],[],{"_220":3629,"_24":236,"_237":3630,"_239":3631},"4a35c5e079f4",[],"F determines which observations become persistent state and how they revise existing records. R selects the state that reaches the next model call. Both can combine stochastic compute and deterministic code. Outcomes may arrive much later than actions. The write path must associate delayed evidence with the original action and revise any lessons derived from it.","631992c5db66",[3635],[],{"_220":3636,"_24":236,"_237":3637,"_239":3638},"904ce784cc6b",[],")","33e2a8d52b3b",[3642],[],{"_220":3643,"_24":236,"_237":3644,"_239":3645},"7680b5ad5684",[],"outcome_t","f4d41d0ffe96",[3649],[],{"_220":3650,"_24":236,"_237":3651,"_239":3652},"80d216e10ccb",[],"action_t,","7f8cc217016d",[3656],[],{"_220":3657,"_24":236,"_237":3658,"_239":3659},"9d5241992604",[],"observation_t,","79b91869e2e3",[3663],[],{"_220":3664,"_24":236,"_237":3665,"_239":3666},"76a995443776",[],"M_t,","f082f460dd34",[3670],[],{"_220":3671,"_24":236,"_237":3672,"_239":3673},"66125ac80fbe",[],"M_t+1 = F(","d3f764a17d9c",[3677],[],{"_220":3678,"_24":236,"_237":3679,"_239":3680},"72a3b7698b71",[],"","cbc2ed7ad74a",[3684],[],{"_220":3685,"_24":236,"_237":3686,"_239":3638},"2816fd2f0683",[],"e6879dfd579c",[3690],[],{"_220":3691,"_24":236,"_237":3692,"_239":3693},"2611c7cb1e9f",[],"observation_t","f22ddf871918",[3697],[],{"_220":3698,"_24":236,"_237":3699,"_239":3700},"affdd49b21ac",[],"context_t,","0b328b527a81",[3704],[],{"_220":3705,"_24":236,"_237":3706,"_239":3707},"2a217f0d1f95",[],"working_state_t,","fbf8e2258ec2",[3711],[],{"_220":3712,"_24":236,"_237":3713,"_239":3714},"ac90fae6dc2e",[],"instructions,","d90d632afed0",[3718],[],{"_220":3719,"_24":236,"_237":3720,"_239":3721},"f9eaeeeb18fd",[],"action_t = model(","bccb7605f45d",[3725],[],{"_220":3726,"_24":236,"_237":3727,"_239":3680},"4b7a59ffa024",[],"62cc0bf02726",[3731],[],{"_220":3732,"_24":236,"_237":3733,"_239":3734},"c20486f8ce97",[],"context_t = R(query_t, M_t, B)","8a7574d4e676",[3738],[],{"_220":3739,"_24":236,"_237":3740,"_239":3741},"6a62e755d8ee",[],"A memory-augmented agent adds persistent state M_t, a write policy F, a retrieval policy R, and a context budget B:","73cf6663ed99",[3745],[],{"_220":3746,"_24":236,"_237":3747,"_239":3748},"28413b799a65",[],"action_t = model(instructions, recent_messages, observation_t)","02f0e6bf9487",[3752],[],{"_220":3753,"_24":236,"_237":3754,"_239":3755},"057f4a37a0a5",[],"At time t, an agent receives an observation, assembles context and chooses an action. We can therefore describe a simplified agent without memory as:","607ef2412da8",[3759],[],{"_220":3760,"_24":236,"_237":3761,"_239":3762},"52f17d0d3b1a",[],"Memory as a Stateful Control Loop","cb4932e369b2",[3766,3767,3768],[],{"_220":3775,"_24":236,"_237":3776,"_239":3777},{"_220":3772,"_24":236,"_237":3773,"_239":3774},{"_220":3769,"_24":236,"_237":3770,"_239":3771},"e314c861bd9f",[]," the model, in the harness layer that controls the agent loop.","d84a78243562",[2077],"outside","778e700cc7c0",[],"It is tempting to describe this as a storage problem. Save the agent's interactions, embed them, and retrieve the closest results before the next model call. That produces persistence, but not necessarily memory. A useful memory system must decide what to write, how to represent it, when to retrieve it, how to resolve conflicting information, and whether/how an outcome should modify future behavior. Current systems implement this functionality ","121f4a444355",[3781],[],{"_220":3782,"_24":236,"_237":3783,"_239":3784},"18e76a78164b",[],"It cannot coordinate with other agents if each action begins with an isolated view of the organization","72565cf8678f",[3788],[],{"_220":3789,"_24":236,"_237":3790,"_239":3791},"744b85306b2c",[],"Learning can't happen without memory, and an agent cannot learn if the outcome of one attempt disappears before the next","a0eeaffd296d",[3795],[],{"_220":3796,"_24":236,"_237":3797,"_239":3798},"0f74cd6c9885",[],"An agent cannot pursue a goal over months if it forgets why the goal/sub-goal matters","7139f92c87c1",[3802],[],{"_220":3803,"_24":236,"_237":3804,"_239":3805},"6a42f6ec679f",[],"To understand agent memory, we shouldn't think of it as a place to store the past. We need to think of it as a harness-level control system that decides which previous experiences should affect future behavior. The need for memory in the long-horizon context should be clear, but to provide a few examples:","51c7e00b5f52",[3812,3813,3814],[3809],{"_220":3810,"_24":1959,"_1960":3811},"f72dcfadb07b","https://www.heavybit.com/library/article/long-horizon-agents-from-order-takers-to-outcome-owners",{"_220":3821,"_24":236,"_237":3822,"_239":3823},{"_220":3818,"_24":236,"_237":3819,"_239":3820},{"_220":3815,"_24":236,"_237":3816,"_239":3817},"ceb09fa7a740",[],", I laid out the five systems I believe are missing before agents can graduate from completing tasks to owning outcomes: Memory, learning, goal orchestration, a durable execution layer, and a toolchain ecosystem. This post dives into agent memory.","d6a782b5390a",[3810],"Part 1","dade4a80a130",[],"In ","829b03b465e4",[3827],[],{"_220":3828,"_24":236,"_237":3829,"_239":3680},"224565565806",[],"16e0361b4ce2",[3833],[],{"_220":3834,"_24":236,"_237":3835,"_239":3836},"991570e6c69f",[],"Why Memory Is A Key Piece of the Long-Horizon Puzzle",[3843,3844,3845,3846,3847,3848,3849,3850,3851,3852,3853,3854,3855,3856,3857,3858,3859,3860,3861,3862,3863,3864,3865,3866,3867,3868,3869,3870,3871,3872,3873,3874,3875,3876,3877,3878,3879,3880,3881,3882,3883,3884,3885,3886,3887,3888,3889,3890,3891,3892,3893],16,"616147c4-aa4e-4ab9-83ae-f20398e6871d","2026-10-20T16:48:00.000Z","how-to-orchestrate-multiple-agents-from-the-terminal","How to Orchestrate Multiple Agents from the Terminal",{"_220":4448,"_24":222,"_223":4449,"_225":4450,"_227":576},{"_220":4430,"_24":222,"_223":4431,"_225":4432,"_227":228},{"_220":4412,"_24":222,"_223":4413,"_225":4414,"_227":228},{"_220":4409,"_24":1973,"_74":4410},{"_220":4395,"_24":222,"_223":4396,"_225":4397,"_227":228},{"_220":4388,"_24":222,"_223":4389,"_225":4390,"_227":576},{"_220":4370,"_24":222,"_223":4371,"_225":4372,"_227":228},{"_220":4347,"_24":222,"_223":4348,"_225":4349,"_227":228},{"_220":4332,"_24":222,"_223":4333,"_225":4334,"_227":228},{"_220":4325,"_24":222,"_223":4326,"_225":4327,"_227":228},{"_220":4318,"_24":222,"_223":4319,"_225":4320,"_227":576},{"_220":4311,"_24":222,"_223":4312,"_225":4313,"_227":228},{"_220":4296,"_24":222,"_223":4297,"_225":4298,"_227":228},{"_220":4289,"_24":222,"_223":4290,"_225":4291,"_227":228},{"_220":4282,"_24":222,"_223":4283,"_225":4284,"_227":228},{"_220":4275,"_24":222,"_223":4276,"_225":4277,"_227":228},{"_220":4268,"_24":222,"_223":4269,"_225":4270,"_227":228},{"_220":4245,"_24":222,"_223":4246,"_225":4247,"_227":228},{"_220":4238,"_24":222,"_223":4239,"_225":4240,"_227":576},{"_220":4231,"_24":222,"_223":4232,"_225":4233,"_227":228},{"_220":4224,"_24":222,"_223":4225,"_225":4226,"_227":228},{"_220":4206,"_24":222,"_223":4207,"_225":4208,"_227":228},{"_220":4199,"_24":222,"_223":4200,"_225":4201,"_227":244},{"_220":4192,"_24":222,"_223":4193,"_225":4194,"_227":576},{"_220":4166,"_24":222,"_223":4167,"_225":4168,"_227":228},{"_220":4151,"_24":222,"_223":4152,"_225":4153,"_227":228},{"_220":4137,"_24":222,"_223":4138,"_225":4139,"_227":228},{"_220":4122,"_24":222,"_223":4123,"_225":4124,"_227":228},{"_220":4115,"_24":222,"_223":4116,"_225":4117,"_227":228},{"_220":4108,"_24":222,"_223":4109,"_225":4110,"_227":228},{"_220":4101,"_24":222,"_223":4102,"_225":4103,"_227":576},{"_220":4086,"_24":222,"_223":4087,"_225":4088,"_227":228},{"_220":4079,"_24":222,"_223":4080,"_225":4081,"_227":228},{"_220":4048,"_24":222,"_223":4049,"_225":4050,"_227":228},{"_220":4041,"_24":222,"_223":4042,"_225":4043,"_227":228},{"_220":4034,"_24":222,"_223":4035,"_225":4036,"_227":576},{"_220":4027,"_24":222,"_223":4028,"_225":4029,"_227":228},{"_220":4020,"_24":222,"_223":4021,"_225":4022,"_227":228},{"_220":4013,"_24":222,"_223":4014,"_225":4015,"_227":228},{"_220":4006,"_24":222,"_223":4007,"_225":4008,"_227":228},{"_220":3999,"_24":222,"_223":4000,"_225":4001,"_227":576},{"_220":3981,"_24":222,"_223":3982,"_225":3983,"_227":228},{"_220":3950,"_24":222,"_223":3951,"_225":3952,"_227":228},{"_220":3943,"_24":222,"_223":3944,"_225":3945,"_227":228},{"_220":3936,"_24":222,"_223":3937,"_225":3938,"_227":228},{"_220":3929,"_24":222,"_223":3930,"_225":3931,"_227":228},{"_220":3922,"_24":222,"_223":3923,"_225":3924,"_227":244},{"_220":3915,"_24":222,"_223":3916,"_225":3917,"_227":576},{"_220":3908,"_24":222,"_223":3909,"_225":3910,"_227":228},{"_220":3901,"_24":222,"_223":3902,"_225":3903,"_227":228},{"_220":3894,"_24":222,"_223":3895,"_225":3896,"_227":228},"3e791b59e928",[3897],[],{"_220":3898,"_24":236,"_237":3899,"_239":3900},"19f013dfac9b",[],"“So I don't think tools like TUIOS go away. I think the job changes. Today it helps you keep an eye on a handful of agents. Next it has to tell you when to look, show you the evidence when you do, and let you jump in at any moment. And honestly, a lot of TUIOS today gets built by agents running inside TUIOS, so I'm motivated to get that part right!”","fb18f8982030",[3904],[],{"_220":3905,"_24":236,"_237":3906,"_239":3907},"a0a3bf5f4ae7",[],"“I think both things happen at the same time. Agents need less babysitting every few months, which means people will run more of them, on longer tasks, across more machines. The total amount of agent work is going to grow way faster than the number of decisions one person can make in a day.”","30122ab66bfb",[3911],[],{"_220":3912,"_24":236,"_237":3913,"_239":3914},"84d5d492af88",[],"Gosain reflects on what will happen to projects like his in the future: Will they become less necessary as individual agents become more competent and easier to manage? Or more necessary as more teams onboard larger swarms of agents?","68a0e61fad4a",[3918],[],{"_220":3919,"_24":236,"_237":3920,"_239":3921},"b2d7e327e929",[],"The Future of Orchestration (From the Terminal or Otherwise)","526f25bd891d",[3925],[],{"_220":3926,"_24":236,"_237":3927,"_239":3928},"e081e6730529",[2077],"“A test that can't fail is worse than no test at all, because it makes you feel safe. I found three tests that had been green for months and could never go red. Now every bug fix needs a test that I've actually watched fail with the fix removed.”","5ce0d3fc1b3e",[3932],[],{"_220":3933,"_24":236,"_237":3934,"_239":3935},"efeec16cdd16",[],"“And listen to your users. Copy mode, custom keybindings, and the daemon all started as issues people opened. Not everyone will like every choice either. Some people think mouse support in a TUI defeats the whole point, and that's fine, keyboard-first is still the default.”","65082acf5730",[3939],[],{"_220":3940,"_24":236,"_237":3941,"_239":3942},"ae3f1436f63e",[],"“Third, write things down. I've written more than 40 posts on the TUIOS blog, mostly about bugs I got wrong the first time. It's made me a better engineer, and I think people trust a project more when it's open about its mistakes.”","83b109315657",[3946],[],{"_220":3947,"_24":236,"_237":3948,"_239":3949},"89189c5dda2c",[],"“Second, figure out early who owns what. A lot of the hardest bugs came from two parts of the system both thinking they owned the same data. One fun one was a full scrollback costing 232MB per pane, and paying it twice, once in the daemon and once in every client.”","6709a19906a8",[3953,3954,3955,3956,3957,3958,3959],[],{"_220":3978,"_24":236,"_237":3979,"_239":3980},{"_220":3975,"_24":236,"_237":3976,"_239":3977},{"_220":3972,"_24":236,"_237":3973,"_239":3974},{"_220":3969,"_24":236,"_237":3970,"_239":3971},{"_220":3966,"_24":236,"_237":3967,"_239":3968},{"_220":3963,"_24":236,"_237":3964,"_239":3965},{"_220":3960,"_24":236,"_237":3961,"_239":3962},"03ff68661418",[]," with the fix removed.”","71d9a0844a69",[2077],"that I've actually watched fail","e821897e33d0",[]," than no test at all, because it makes you feel safe. I found three tests that had been green for months and could never go red. Now every bug fix needs a test ","9fcb2fef5254",[2077],"worse","89f7daac29b3",[]," is ","c68efdd5d660",[2077],"a test that can't fail","e7a2992b03e7",[],"“The other half of that lesson is that ","2c567aa855f5",[3987,3988,3989],[3984],{"_220":3985,"_24":1959,"_1960":3986},"f924a1482c30","https://github.com/Gaurav-Gosain/tuitest",{"_220":3996,"_24":236,"_237":3997,"_239":3998},{"_220":3993,"_24":236,"_237":3994,"_239":3995},{"_220":3990,"_24":236,"_237":3991,"_239":3992},"807d067e16ed",[],".”","e6b851d76bb4",[3985],"a separate tool for it","1df3f5508e3e",[],"The creator shares many important lessons learned while building the project. “Testing is the big one. Early on I literally wrote in an issue that I was still trying to figure out how to test TUIOS properly, because testing a TUI with real shells inside it is way harder than it looks. Now the tests I trust most start the real binary, attach a real client and read what's on the screen. I even ended up building ","52e66ff87303",[4002],[],{"_220":4003,"_24":236,"_237":4004,"_239":4005},"b09cc780d360",[],"Lessons Learned: Testing, Ownership, Docs, Community","28c61fd9c130",[4009],[],{"_220":4010,"_24":236,"_237":4011,"_239":4012},"f03d323caae1",[],"“For bigger organizations, I'd expect higher-level tools on top that handle goals, ownership and cost, and TUIOS is built to sit underneath those. Pretty much everything it does is available as a command, as JSON, as an event stream and through MCP, so something else can drive it while a person can still jump into any agent whenever they want.”","bd455af80937",[4016],[],{"_220":4017,"_24":236,"_237":4018,"_239":4019},"731f4a1033a5",[],"For long-horizon runs, the creator points out that sessions require a different layer underneath that supports persistent sessions beyond a laptop close. With a significant amount of time and effort, the creator was eventually able to implement a much-requested daemon to restore sessions post-restart into this project, but larger orgs will presumably need more.","1b8cce8046c1",[4023],[],{"_220":4024,"_24":236,"_237":4025,"_239":4026},"9c0a305c08ff",[],"“When an agent gets stuck I want to see exactly what it sees, type into its session, and take over if I need to. That's really natural in a terminal and pretty awkward from a dashboard. It's also where agents already do their work anyway, with the shell, git and tests.”","ffdc999356dc",[4030],[],{"_220":4031,"_24":236,"_237":4032,"_239":4033},"302606d775e7",[],"While the terminal remains popular for individual devs, Gosain isn’t sure it will scale for major projects across an entire organization. “I think the terminal is the right place for agents to run and for you to step in, but not where you'd set goals and budgets for a whole department.\"","f3e6dbfc6664",[4037],[],{"_220":4038,"_24":236,"_237":4039,"_239":4040},"3a4e64e47a2d",[],"The Future of the Terminal: Long Horizon and Org-Wide Projects","3cc74079e800",[4044],[],{"_220":4045,"_24":236,"_237":4046,"_239":4047},"cf20c36daa7e",[],"Gosain adds that untrusted text remains a potential loophole, as agents will read anything from anywhere and potentially interpret what they read as instructions. Text hidden in PR titles, invisible characters in Unicode, and others. “That kind of labelling really needs to become standard everywhere.”","0dfa637b8b71",[4051,4052,4053,4054,4055,4056,4057],[],{"_220":4076,"_24":236,"_237":4077,"_239":4078},{"_220":4073,"_24":236,"_237":4074,"_239":4075},{"_220":4070,"_24":236,"_237":4071,"_239":4072},{"_220":4067,"_24":236,"_237":4068,"_239":4069},{"_220":4064,"_24":236,"_237":4065,"_239":4066},{"_220":4061,"_24":236,"_237":4062,"_239":4063},{"_220":4058,"_24":236,"_237":4059,"_239":4060},"255ac56e2d27",[],". You should be able to take it away while the agent is still running. And you need a clear record of which agent did what, and on whose behalf.”","95994052226c",[2077],"more","47dfd88b7b4f",[]," than it has but never ","1a9fe7c73af4",[2077],"less","860c3089b9db",[],", so an agent can give a helper ","837e89a39a8a",[2077],"down","ec06fc46f793",[],"“Building grants taught me a few things agents need that most systems don't have yet. Access should be tied to a task rather than a person. It should only ever be passed ","cdb499682691",[4082],[],{"_220":4083,"_24":236,"_237":4084,"_239":4085},"7512596d2d33",[],"The creator explains that while his project focused on least privilege rather than full-blown identity management, teams will likely need IAM systems for both human engineers and agents in the future (and major identity vendors are already hard at work building them).","45b541c5899f",[4089,4090,4091],[],{"_220":4098,"_24":236,"_237":4099,"_239":4100},{"_220":4095,"_24":236,"_237":4096,"_239":4097},{"_220":4092,"_24":236,"_237":4093,"_239":4094},"96fa96fae98a",[],". “Every pane in TUIOS has grants that say what the program inside it is allowed to do through TUIOS: Read its own session, type into it, start other agents, answer prompts, or do everything. The daemon checks every single call against them, no matter how it comes in. So if I want a reviewer agent that can look but not touch, I just start it with read-only.”","2083e36a4d74",[2077],"specifically run through the project itself","c439e261144e",[],"Gosain points out that the project has built-in identity management for jobs ","bd5192f93adf",[4104],[],{"_220":4105,"_24":236,"_237":4106,"_239":4107},"bf4caaeb1172",[],"Managing Agents Via Terminal: Identity and Loopholes","df24faceca60",[4111],[],{"_220":4112,"_24":236,"_237":4113,"_239":4114},"29495e2c94ed",[],"The creator also points out that his project has its limits, controlling what an agent can do through its multiplexer, but not everywhere else. “The ‘ask a human’ path has to be quick and painless, otherwise people just turn it off.”","ed32ddd9d6d8",[4118],[],{"_220":4119,"_24":236,"_237":4120,"_239":4121},"a346f352c481",[],"Gosain suggests that agents generally don’t go astray because they’re inherently malicious. “They go wrong because they're confident and fast. We've all seen the stories of an agent wiping a production database because it found a token lying around. That's what I worry about, an agent being completely sure about something it really should have asked about.”","7918883b39e8",[4125,4126,4127],[],{"_220":4134,"_24":236,"_237":4135,"_239":4136},{"_220":4131,"_24":236,"_237":4132,"_239":4133},{"_220":4128,"_24":236,"_237":4129,"_239":4130},"66999ff3e33e",[]," a risky request if you've allowed it to, but it can never approve one. An agent can't give itself more access, and it can't give a helper more access than it has. Even if an agent edits the config file to give itself more, nothing changes until you apply that change yourself from outside TUIOS.”","dbe58bfdd53e",[2077],"deny","9c3818dff01a",[],"“In practice, that means only you can answer an agent's permission prompt. Another agent can ","d4a501c3e7f1",[4140,4141,4142],[],{"_220":4148,"_24":236,"_237":4149,"_239":4150},{"_220":4145,"_24":236,"_237":4146,"_239":4147},{"_220":4143,"_24":236,"_237":4144,"_239":3992},"79b7b12169fb",[],"fb8f26544685",[2077],"something risky","1bd9d2a22783",[],"According to Gosain, learning from agents on the job is valuable, but only within the bounds of limited autonomy. “The rough rule I ended up with is that agents can do the work and agents can say no, but only a person gets to say yes to ","60319847686f",[4154,4155,4156],[],{"_220":4163,"_24":236,"_237":4164,"_239":4165},{"_220":4160,"_24":236,"_237":4161,"_239":4162},{"_220":4157,"_24":236,"_237":4158,"_239":4159},"729256caecec",[],".” The creator explains that fans let you run your own checks on every attempt, compare the results side by side, and diff the attempts against each other. To help users improve on their own processes, the project allows for annotation that can be shared with agents, effectively giving users the final say based on specific evidence pulled from each agentic run.","715fe426e713",[2077],"comparing","ba5da87edf7b",[],"“The part I care about most is ","f65d4a7bf402",[4172,4173,4174,4175,4176],[4169],{"_220":4170,"_24":1959,"_1960":4171},"3ed3816cd07a","https://github.blog/ai-and-ml/github-copilot/what-are-git-worktrees-and-why-should-i-use-them/",{"_220":4189,"_24":236,"_237":4190,"_239":4191},{"_220":4186,"_24":236,"_237":4187,"_239":4188},{"_220":4183,"_24":236,"_237":4184,"_239":4185},{"_220":4180,"_24":236,"_237":4181,"_239":4182},{"_220":4177,"_24":236,"_237":4178,"_239":4179},"20732e20c65c",[],", partly as a learning aid, but mostly to compare approaches. “Different models fail...and succeed in different ways. If I give the same problem to three agents, I'll often get three reasonable approaches, and seeing them side by side helps me think about the problem. Agent time is cheap and mine isn't, so it's a pretty good trade.”","33fc1103c318",[4170],"worktree","d13353fe843a",[]," functionality to parallelize task execution across multiple agents, each in its own ","46ba47a14fde",[2077],"fan","ca141d2fe81c",[],"The creator notes that he specifically built a ","8fc105268f08",[4195],[],{"_220":4196,"_24":236,"_237":4197,"_239":4198},"291bd897e642",[],"Failstates, Agent Reliability, and Learning Job-to-Job","80403ff8cd0a",[4202],[],{"_220":4203,"_24":236,"_237":4204,"_239":4205},"8d2aebf3369f",[2077],"“Pretty much every lab shipped its own ‘which of my agents needs me’ view, and each one only knows about its own agent. That's fine if you only use one tool, but most people I know don't.” - Gaurav Gosain/Creator, TUIOS","3dca538bee60",[4212,4213,4214],[4209],{"_220":4210,"_24":1959,"_1960":4211},"eb4ac0b4ccd2","https://www.cnbc.com/2026/06/16/anthropics-fable-shutdown-is-a-big-moment-for-open-source-ai.html",{"_220":4221,"_24":236,"_237":4222,"_239":4223},{"_220":4218,"_24":236,"_237":4219,"_239":4220},{"_220":4215,"_24":236,"_237":4216,"_239":4217},"db1ed4e125ad",[]," [access to its tools].” Rather than depend on any one vendor’s roadmap, the project was built to work with multiple tools as well as to be directly compatible with tmux through a tmux shim.","af462a24bae4",[4210],"a lab decided to cut off","63e88dfb40d6",[],"The creator clarifies that his goal wasn’t to create a defensible moat around a single, winner-take-all harness that outcompetes all other labs’ harnesses. “[This] is exactly why a neutral layer is useful. We also saw how fast access can change, when ","c330be734330",[4227],[],{"_220":4228,"_24":236,"_237":4229,"_239":4230},"8217051e8152",[],"“Something I noticed this year is that pretty much every lab shipped its own ‘which of my agents needs me’ view, and each one only knows about its own agent. That's fine if you only use one tool, but most people I know don't. TUIOS recognizes 24 agent CLIs right now, and anything it doesn't know still runs fine as a normal terminal program.”","ab4bc0e3ed05",[4234],[],{"_220":4235,"_24":236,"_237":4236,"_239":4237},"b84cc79e55a3",[],"Gosain is frank about why he decided to keep the project as something that sits on top of other harnesses, rather than trying to also build his own coding agent. “The harnesses get better every week, and the teams behind them are a lot bigger than one person working on an open source project on the side. I also just don't think another agent is what people need. Most people have their favorite, and a lot of them use two or three depending on the task.”","db2551a7b932",[4241],[],{"_220":4242,"_24":236,"_237":4243,"_239":4244},"a7aeb5be88b4",[],"Where Terminal-Level Orchestration Fits in the Big Picture","fa331092dd99",[4248,4249,4250,4251,4252],[],{"_220":4265,"_24":236,"_237":4266,"_239":4267},{"_220":4262,"_24":236,"_237":4263,"_239":4264},{"_220":4259,"_24":236,"_237":4260,"_239":4261},{"_220":4256,"_24":236,"_237":4257,"_239":4258},{"_220":4253,"_24":236,"_237":4254,"_239":4255},"6d9a9be4822a",[]," won't, at least not soon. So review doesn't disappear, it shifts from reading every line to looking at evidence, like whether the tests pass or how two attempts compare.”","03e18c2bea33",[2077],"‘Is this actually the right change?’","14d1aa383e6c",[]," will mostly get automated away. ","459f8e217e2c",[2077],"‘Can this agent run this command?’","f60396b582a3",[],"“I also think the bottleneck moves over time. ","cedcc0d17912",[4271],[],{"_220":4272,"_24":236,"_237":4273,"_239":4274},"472e63a58ba6",[],"The creator suggests that his approach is about streamlining the decision process to fewer, better questions with the assumption that the routine stuff can and will run on its own. “And when something does reach you, it comes with enough context that you can decide in a few seconds.”","c8802a545e46",[4278],[],{"_220":4279,"_24":236,"_237":4280,"_239":4281},"c3b40f8ab7f9",[],"“But I also don't think the answer is asking the human about everything. If you get asked a hundred times an hour, you stop reading and just hit yes.” Gosain suggests that major AI labs are trending in the direction of letting classifiers handle routine approvals, which is probably for the best.","ecf4076fb794",[4285],[],{"_220":4286,"_24":236,"_237":4287,"_239":4288},"dde60bb1e02b",[],"The creator suggests that the bottleneck in human-plus-agent systems is still, from his perspective, human attention and decision-making. “I think that's how it should be. I don't want an agent deciding on its own to force push or delete something important.","4ea2d7f44903",[4292],[],{"_220":4293,"_24":236,"_237":4294,"_239":4295},"96ed4d810536",[],"The creator suggests that the project evolved to help human users spend their limited attention span more wisely. “Agents report what state they're in and the ones that need you [get floated] to the top.” The project collects questions and approvals, and agent turns get saved at git checkpoints to make fixes easier, with cheap verification being something to work on in the future.","25ebcace03e7",[4299,4300,4301],[],{"_220":4308,"_24":236,"_237":4309,"_239":4310},{"_220":4305,"_24":236,"_237":4306,"_239":4307},{"_220":4302,"_24":236,"_237":4303,"_239":4304},"a9b7b4c3a230",[],". Spinning up agents is the easy part now. Every lab ships some way to run them in parallel, and some can kick off hundreds of subagents in one go. But producing a hundred results is cheap, and figuring out which ten are actually good is not. If you can't review it, running more of it doesn't really help you.”","c9a437749d53",[2077],"attention","e98c0abcd22c",[],"“What breaks first, at least for me, is ","7b661806cc77",[4314],[],{"_220":4315,"_24":236,"_237":4316,"_239":4317},"1f0ef4e47a71",[],"Gosain suggests that there’s more to any agentic job than simply how many agents are in play. “I don't think there's a linear curve like five, then 10, then 100 [agents being a breaking point]. It depends on the work. Ten agents doing well-defined changes against a good test suite is honestly easier to manage than two agents doing open-ended design work.”","3f31d5748e7c",[4321],[],{"_220":4322,"_24":236,"_237":4323,"_239":4324},"467d878226a1",[],"Where Do Agent-Plus-Human-in-the-Loop Systems Break?","724b27bc87e5",[4328],[],{"_220":4329,"_24":236,"_237":4330,"_239":4331},"8a8f5b8994d7",[],"“TUIOS already owned every one of those panes, so it felt obvious that it should be the one keeping track.” For the creator, giving cues to help agentic power users like himself was becoming increasingly important (though the project still offers an ‘agents-off’ feature for those who just want a better multiplexer).","23ee3f5a58b7",[4335,4336,4337],[],{"_220":4344,"_24":236,"_237":4345,"_239":4346},{"_220":4341,"_24":236,"_237":4342,"_239":4343},{"_220":4338,"_24":236,"_237":4339,"_239":4340},"5382a94334d1",[]," became the problem. “An agent would hit a permission prompt, draw it once, and then just sit there.” Different screens that were in different project states looked identical at a glance, and would sit waiting for a human thumbs up. “It was a pretty silly way to lose time.”","56cdea4cbb7d",[2077],"which agent needed him","01c30f4c4a0b",[],"As Gosain started running several agents in parallel, arranging panes stopped being the problem. Knowing ","c8f8b537e2f0",[4350,4351,4352,4353,4354],[],{"_220":4367,"_24":236,"_237":4368,"_239":4369},{"_220":4364,"_24":236,"_237":4365,"_239":4366},{"_220":4361,"_24":236,"_237":4362,"_239":4363},{"_220":4358,"_24":236,"_237":4359,"_239":4360},{"_220":4355,"_24":236,"_237":4356,"_239":4357},"bcfd2c3fa106",[],"”","b4a8ddc05174",[2077],"which one of them was waiting on me.","f8ecbb456c9b",[]," became the part I actually cared about. I didn't need help arranging my windows anymore, I needed something to tell me ","f55848acbc67",[2077],"that","52a26e91295d",[],"“A multiplexer sees every program you run, its output, its title, the folder it's in, and yet it never really did anything with any of that. Once I started running coding agents in my panes, ","2c0547539626",[4376,4377,4378],[4373],{"_220":4374,"_24":1959,"_1960":4375},"66e4fafd0d2e","https://github.com/tmux/tmux/wiki",{"_220":4385,"_24":236,"_237":4386,"_239":4387},{"_220":4382,"_24":236,"_237":4383,"_239":4384},{"_220":4379,"_24":236,"_237":4380,"_239":4381},"d57416840743",[]," is great, it always felt stuck in a world of memorized key chords and config files.” The creator points out that while the project was initially about all-in-one window management UX, it soon highlighted a deeper need.","c328590f47fd",[4374],"tmux","b55fff7bf397",[],"Gosain notes that the project started from pursuing a better terminal experience. “I live in the terminal for pretty much everything, and while ","3a705bf2cb04",[4391],[],{"_220":4392,"_24":236,"_237":4393,"_239":4394},"49637b6e2cd9",[],"From Multiple Terminals to Multi-Agent","98f9ba5b1089",[4401,4402],[4398],{"_220":4399,"_24":1959,"_1960":4400},"f04212d3515e","https://www.tuios.dev",{"_220":4406,"_24":236,"_237":4407,"_239":4408},{"_220":4403,"_24":236,"_237":4404,"_239":4405},"9b006e84ff8b",[4399,2077],"tuios.dev","51417f32ca16",[2077],"TUIOS helps users orchestrate agents across multiple terminals. Image courtesy ","d4790a7a95ba",{"_1944":4411,"_24":1946},"image-31b56f7b3c41c15c73d7b410e290c2f05c7c5ede-1600x478-jpg","fa72a247de6c",[4418,4419,4420],[4415],{"_220":4416,"_24":1959,"_1960":4417},"3e20d7e69a4e","https://github.com/Gaurav-Gosain/tuios",{"_220":4427,"_24":236,"_237":4428,"_239":4429},{"_220":4424,"_24":236,"_237":4425,"_239":4426},{"_220":4421,"_24":236,"_237":4422,"_239":4423},"d4349a853895",[]," in response. On the surface, TUIOS acts as a multiterminal orchestration tool, but as he explains below, it’s ultimately a way to address one of the trickiest new bottlenecks in software: The limited attention of a human user.","7ed88089c064",[4416],"TUIOS","50d2ad72640b",[],"But what happens when you start utilizing so many agents at once, and have to have so many terminal windows open that it becomes difficult to keep track of everything? Open-source creator Gaurav Gosain noticed this limitation and began building his project ","5dad60941206",[4436,4437,4438],[4433],{"_220":4434,"_24":1959,"_1960":4435},"52e334b19f82","https://en.wikipedia.org/wiki/Terminal_multiplexer",{"_220":4445,"_24":236,"_237":4446,"_239":4447},{"_220":4442,"_24":236,"_237":4443,"_239":4444},{"_220":4439,"_24":236,"_237":4440,"_239":4441},"932af58ea732",[]," to give each agent their own separate pane that segregates and preserves the work of each across sessions.","ae81e2f7bd1b",[4434],"terminal multiplexers","fcf5bef05f52",[],"With the increasing popularity of command-line interface (CLI) building for agentic projects, terminal interfaces are seeing something of a rebirth in software development. However, as more developers utilize agentic tools, and increasingly, multiple agents at once, they’re turning to ","eef26fcaf517",[4451],[],{"_220":4452,"_24":236,"_237":4453,"_239":4454},"24ae57392a2a",[],"Do Windows Matter Less Than Limited Human Attention?",{"_2000":4488,"_96":4489,"_1973":4490,"_2004":1992,"_4461":4462,"_51":4491,"_14":4492,"_63":4493,"_2011":25},{"_2000":4474,"_96":4475,"_1973":4476,"_2004":1990,"_4461":4462,"_51":4477,"_14":4478,"_63":4479,"_2011":25},{"_2000":4458,"_96":4459,"_1973":4460,"_2004":1988,"_4461":4462,"_51":4463,"_14":4464,"_63":4465,"_2011":25},[4466],859.0021,"https://image.mux.com/ZfDsqm02Bjw4U1qiGt2XcKYLXgQo4sdL02/thumbnail.png","poster","mux","2026-10-26T15:25:00.000Z","safety-for-autonomous-organizations","Safety for Autonomous Organizations",{"_220":4467,"_24":222,"_223":4468,"_225":4469,"_227":228},"exsafe0",[4470],[],{"_220":4471,"_24":236,"_237":4472,"_239":4473},"exsafe0s",[],"Axel Backlund, cofounder of Andon Labs and creator of Vending-Bench, shares what agents running vending machines, a market, a cafe, and a radio station reveal about long-horizon autonomy, and why models that can tell when they are in a simulation make real-world deployments essential to safety work.",[4480],1381.3823,"https://image.mux.com/v1FWaHcoYP7EpNyiNbxkyvFqG02JeSzx7/thumbnail.png","2026-10-26T15:30:00.000Z","performance-engineering-for-agents-lessons-from-glean","Performance Engineering for Agents: Lessons from Glean",{"_220":4481,"_24":222,"_223":4482,"_225":4483,"_227":228},"exglean0",[4484],[],{"_220":4485,"_24":236,"_237":4486,"_239":4487},"exglean0s",[],"Tony Gentilcore, cofounder of Glean and a veteran of Chrome's Speed Team, brings a performance-first lens to long-horizon agents in a fireside with Chad Metcalf: reading traces, token budgets that change agent behavior, what work belongs in the background, and measuring spend per unit of throughput.",[4494],1056.5995,"https://image.mux.com/4jMAtMC7uNlI01kBXJ00i6R00SWSPhae005l/thumbnail.png","2026-10-26T15:35:00.000Z","the-mythical-agent-month","The Mythical Agent-Month",{"_220":4495,"_24":222,"_223":4496,"_225":4497,"_227":228},"exwes0",[4498],[],{"_220":4499,"_24":236,"_237":4500,"_239":4501},"exwes0s",[],"Wes McKinney, cofounder of Kenn Software, creator of pandas, and co-creator of Apache Arrow, revisits The Mythical Man-Month for the age of agent factories: why coordination is part of the work, and why judgment and taste matter more as agents take on more of the building.",{"_220":4503,"_24":222,"_223":4504,"_225":4505,"_227":228},"exgoals0",[4506],[],{"_220":4507,"_24":236,"_237":4508,"_239":71},"exgoals0s",[],"ua","Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; ClaudeBot/1.0; +claudebot@anthropic.com)","hideBanner",{"_4531":4532,"_4533":4534,"_4535":4536,"_4537":-5},"ENV",{"_4515":4516,"_4517":4518,"_4519":4520,"_4521":4522,"_4523":4524,"_4525":4526,"_4527":4528,"_4529":4516,"_4530":4516},"SANITY_DATASET","production","ALGOLIA_INDEX","heavybit-library","ALGOLIA_APPLICATION_ID","0QCKBC1FMK","ALGOLIA_SITEWIDE_INDEX","heavybit-sitewide","ALGOLIA_JOBS_INDEX","heavybit-jobs","MUX_ENV_KEY","dRI98bL8ze03gYDL6-Ogle2Js","ALGOLIA_SEARCH_API_KEY","7733fdd199ec7af7ee2923ec39c3f036","MODE","VITE_CONTEXT","footer",{"_4649":4650,"_4651":4652,"_4653":4654},"globalConfig",{"_72":4646},"header",{"_4538":4539,"_4540":4541},"skills","announcementBar",{"_4638":4639,"_4640":-5,"_4641":31,"_4642":4643,"_239":4644,"_90":4645},"headerNavLinks",[4542,4543,4544,4545,4546],{"_220":4619,"_4548":4620,"_4549":4621,"_1959":4622,"_4552":12,"_63":4623},{"_220":4604,"_4548":4605,"_4549":4606,"_1959":4607,"_4552":12,"_63":4608},{"_220":4586,"_4548":4587,"_4549":4588,"_1959":4589,"_4552":12,"_63":4590},{"_220":4554,"_4548":4555,"_4549":4556,"_1959":4557,"_4552":12,"_63":4558},{"_220":4547,"_4548":-5,"_4549":4550,"_1959":4551,"_4552":31,"_63":4553},"3702498a502c","dropdown","label","Contact","/press","mobileOnly","Navigate to contact page","4bec0dd665d4",[4559,4560,4561,4562,4563,4564],"Library","/library","Navigate to library page",{"_220":4582,"_4549":4583,"_1959":4584,"_63":4585},{"_220":4578,"_4549":4579,"_1959":4580,"_63":4581},{"_220":4574,"_4549":4575,"_1959":4576,"_63":4577},{"_220":4571,"_4549":4572,"_1959":4573,"_63":4572},{"_220":4568,"_4549":4569,"_1959":4570,"_63":4569},{"_220":4565,"_4549":4566,"_1959":4567,"_63":4566},"66773779036c","DevToolsDigest","/devtoolsdigest","2ead3bd14715","Reports","/reports","ac06751f0372","Collections","/library/collections","1f3fb936f3b9","Podcast Network","/library/podcasts","Navigate to podcasts page","6526c69f5260","Videos","/library?filter=videos","Watch educational videos in Heavybit's Library","63e48d609548","Articles","/library?filter=articles","Read educational articles in Heavybit's Library","716b76469486",[4591,4592,4593],"Community","/community","Navigate to community page",{"_220":4601,"_4549":4602,"_1959":4603,"_63":4602},{"_220":4598,"_4549":4599,"_1959":4600,"_63":4599},{"_220":4594,"_4549":4595,"_1959":4596,"_63":4597},"23077fca0e2f","DevGuild","/devguild","Navigate to devguild page","1fb7c9c465d0","Advisors","/community/advisors","3a177efc4649","Events","/events","90b68b861c06",[4609,4610],"Portfolio","/portfolio","Navigate to portfolio page",{"_220":4615,"_4549":4616,"_1959":4617,"_63":4618},{"_220":4611,"_4549":4612,"_1959":4613,"_63":4614},"e50fa43ad2e7","Jobs","/jobs","Portfolio Company Jobs","d481fcbbd3af","Spotlights","/portfolio/spotlights","Navigate to spotlights page","a8d641babed2",[4624,4625,4626],"About","/about","Navigate to about page",{"_220":4634,"_4549":4635,"_1959":4636,"_63":4637},{"_220":4630,"_4549":4631,"_1959":4632,"_63":4633},{"_220":4627,"_4549":4628,"_1959":4551,"_63":4629},"84e68f9f5c33","News","Navigate to press page","3d279775d73b","FAQ","/faq","Navigate to FAQ page","964627410473","Team","/team","Navigate to team page","buttonText","RSVP","end","show","start","2026-10-20T21:46:59.960Z","🐺 [Oct 28] GitBlame: Werewolf Happy Hour @GitHub Universe","https://luma.com/gitblame-werewolf",{"_24":1973,"_74":4647},{"_1944":4648,"_24":1946},"image-32fdfae4fd13b8f9a3c5bf6b13a635068cb0ecbb-1332x529-png","emailSubscription",{"_1879":4720,"_4721":4722},"footerColumns",[4686,4687,4688],"socialMedia",[4655,4656,4657,4658,4659],{"_220":4681,"_4661":4682,"_90":4683},{"_220":4676,"_4661":4677,"_90":4678},{"_220":4671,"_4661":4672,"_90":4673},{"_220":4666,"_4661":4667,"_90":4668},{"_220":4660,"_4661":4662,"_90":4663},"639729864c2d","icon",{"_24":1942,"_74":4664},"https://www.tiktok.com/@heavybitindustries",{"_1944":4665,"_24":1946},"image-48ac9473e30341c4bee9f6634b77b6f8f1dd6b2f-849x980-webp","a100f6448036",{"_24":1942,"_74":4669},"https://www.instagram.com/heavybitindustries/",{"_1944":4670,"_24":1946},"image-b38b7ee9ccd05f61ca43d5b29fe7058f822f87eb-512x512-webp","fae1cc0df5c6",{"_24":1942,"_74":4674},"https://www.youtube.com/c/heavybit",{"_1944":4675,"_24":1946},"image-6c548d2e815d1fb3b3077c9b8be1ec251e1bc418-512x512-png","5348b028f0d41c3dac62b3dd2b5531ee",{"_24":1942,"_74":4679},"https://www.linkedin.com/company/heavybit",{"_1944":4680,"_24":1946},"image-01bd92eb06873bdc30b1cb5f9c78b1ef5a9559d4-512x512-png","ee025c96397c",{"_24":1942,"_74":4684},"https://twitter.com/heavybit",{"_1944":4685,"_24":1946},"image-9d618da0948299428b29e5d499af3fa5369a24d2-20x20-svg",{"_220":4712,"_4690":4621,"_1993":4713},{"_220":4700,"_4690":4588,"_1993":4701},{"_220":4689,"_4690":4556,"_1993":4691},"ea618aa34610","heading",[4692,4693,4694],{"_220":4697,"_4549":4698,"_1959":4557,"_63":4699},{"_220":4696,"_4549":4575,"_1959":4576,"_63":4577},{"_220":4695,"_4549":4566,"_1959":4567,"_63":4566},"96a97ceba3a2","e1673433290e","c82a64edc4dd","Blog","Navigate to browse on library page","a584a6acf07b",[4702,4703,4704,4705],{"_220":4711,"_4549":4602,"_1959":4603,"_63":4602},{"_220":4710,"_4549":4595,"_1959":4596,"_63":4597},{"_220":4709,"_4549":4612,"_1959":4613,"_63":4614},{"_220":4706,"_4549":4707,"_1959":4708,"_63":4707},"e22bfa6b2330","Inclusion","/community/diversity-and-inclusion","feb038ce46a1","e2d5641ef7de","4358699f4ca4","30a87ef6e7c1",[4714,4715,4716],{"_220":4719,"_4549":4606,"_1959":4607,"_63":4608},{"_220":4718,"_4549":4635,"_1959":4636,"_63":4637},{"_220":4717,"_4549":4628,"_1959":4551,"_63":4629},"398a9b50e214","42666d0b3987","3dfef1283ab5","Unlock success from technical pioneers in software infrastructure. We send the best strategies, tools, and industry updates directly to your inbox.","targetURL","https://share.hsforms.com/1_KYaR4lAQm-C8_9H7r4MIwqdoyu"]

Goals! Orchestration and the Human Control Plane

Speakers