[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8},"actionData","errors","root",{"_3719":3720,"_3721":-5,"_9":3722,"_3723":3724},"routes/library.video.$slug",{"_9":10,"_11":12,"_13":-5,"_14":15},"data",{"_16":17},"preview",false,"pageNumber","slug","the-long-horizon-stack","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","18b72251-7b14-45b2-ab6c-ca370ea48cd5","_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",[3712],"filmedOn","2026-09-24","library",{"_1205":1206,"_1207":1208},"modules","moreFromLibrary",[1196,1197,1198],"outlines",[1188],"posterFrame",{"_24":1091,"_74":1186},"publishDate","2026-10-26T15:00:00.000Z","skill","slideDeck",{"_24":14,"_984":15},"speakers","speakersWithOrg",[1026],"startDate","2026-09-24T16:30:00.000Z","tags",[968,969,970,971,972,973,974],"title","The Long-Horizon Stack","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],{"_24":73,"_74":75},"seo",{"_70":71,"_72":50},"OGdescription","Heavybit General Partner Amir Zohrenejad opens DevGuild with the case that long horizon agents now work, citing OpenAI's 10,000-agent, 88-hour run at the Navier-Stokes problem. He then names the five systems that still need to mature before most engineering teams can deploy them: memory, goal orchestration, durable execution, learning, and tools.","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},"2026-10-23T21:58:29Z","a1ced262-8657-44e1-8545-5d7566ade8c1","YEARsfayN2fPM1bkAoF3JA","mux.videoAsset","2026-10-23T22:00:53Z","assetId","01402nGiv4DUKgOULMyk1omhhuxV402jZ9t",{"_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","nF9m8c3G1rGAPEHjwbgA8kar5Z29FABx","status","ready","uploadId","iCUCYaeeau1IN4n3RzeJrfA1DS5g5yMX","url","https://stream.mux.com/nF9m8c3G1rGAPEHjwbgA8kar5Z29FABx","aspect_ratio","16:9","created_at","1792792725","duration",477.387533,"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},477.343533,"3SACHX1WIKmjxx6DuxwqWIaum1SeHTZYfEiREXp9jmQ","max_frame_rate","max_height",1152,"max_width",2048,"type",477.343542,"tACUevEXoQ01U00Cbr6lroHFXdjXIqATAT","max_channel_layout","stereo","max_channels",2,"primary","audio",100,"state","completed",{"_100":85,"_147":148},"policy","public","video_resolution","3840x2160",{"_186":961,"_24":188,"_189":962,"_191":963,"_193":422},{"_186":941,"_24":188,"_189":942,"_191":943,"_193":221},{"_186":921,"_24":188,"_189":922,"_191":923,"_193":221},{"_186":901,"_24":188,"_189":902,"_191":903,"_193":221},{"_186":881,"_24":188,"_189":882,"_191":883,"_193":221},{"_186":861,"_24":188,"_189":862,"_191":863,"_193":221},{"_186":831,"_24":188,"_189":832,"_191":833,"_193":221},{"_186":811,"_24":188,"_189":812,"_191":813,"_193":221},{"_186":791,"_24":188,"_189":792,"_191":793,"_193":221},{"_186":751,"_24":188,"_189":752,"_191":753,"_193":221},{"_186":731,"_24":188,"_189":732,"_191":733,"_193":221},{"_186":711,"_24":188,"_189":712,"_191":713,"_193":194},{"_186":701,"_24":188,"_189":702,"_191":703,"_193":221},{"_186":694,"_24":188,"_189":695,"_191":696,"_193":422},{"_186":664,"_24":188,"_189":665,"_191":666,"_193":221},{"_186":634,"_24":188,"_189":635,"_191":636,"_193":221},{"_186":614,"_24":188,"_189":615,"_191":616,"_193":221},{"_186":594,"_24":188,"_189":595,"_191":596,"_193":221},{"_186":574,"_24":188,"_189":575,"_191":576,"_193":221},{"_186":534,"_24":188,"_189":535,"_191":536,"_193":221},{"_186":504,"_24":188,"_189":505,"_191":506,"_193":221},{"_186":474,"_24":188,"_189":475,"_191":476,"_193":221},{"_186":467,"_24":188,"_189":468,"_191":469,"_193":422},{"_186":447,"_24":188,"_189":448,"_191":449,"_193":221},{"_186":427,"_24":188,"_189":428,"_191":429,"_193":221},{"_186":419,"_24":188,"_189":420,"_191":421,"_193":422},{"_186":389,"_24":188,"_189":390,"_191":391,"_193":221},{"_186":359,"_24":188,"_189":360,"_191":361,"_193":221},{"_186":349,"_24":188,"_189":350,"_191":351,"_193":221},{"_186":319,"_24":188,"_189":320,"_191":321,"_193":221},{"_186":289,"_24":188,"_189":290,"_191":291,"_193":221},{"_186":269,"_24":188,"_189":270,"_191":271,"_193":221},{"_186":249,"_24":188,"_189":250,"_191":251,"_193":221},{"_186":218,"_24":188,"_189":219,"_191":220,"_193":221},{"_186":187,"_24":188,"_189":190,"_191":192,"_193":194},"_key","329fe41e36b2","block","children",[203,204,205],"markDefs",[195,196],"style","blockquote",{"_186":201,"_24":198,"_199":202},{"_186":197,"_24":198,"_199":200},"8ca3a82","timestamp","time","00:07:45.440","bee31a6","00:07:42.720",{"_186":215,"_24":207,"_208":216,"_210":217},{"_186":212,"_24":207,"_208":213,"_210":214},{"_186":206,"_24":207,"_208":209,"_210":211},"1680251f9d8d","span","marks",[197],"text","The learning, the memory, these things have to be shared between the tools and the harness and the model.","1fad642763d9",[]," ","954ff029adc1",[201],"The context has to flow back from the tools.","d3e46e730875",[231,232,233,234,235],[222,223,224],"normal",{"_186":229,"_24":198,"_199":230},{"_186":227,"_24":198,"_199":228},{"_186":225,"_24":198,"_199":226},"97cfd30","00:07:37.240","94d12ae","00:07:31.320","c5023d8","00:07:21.380",{"_186":246,"_24":207,"_208":247,"_210":248},{"_186":244,"_24":207,"_208":245,"_210":214},{"_186":241,"_24":207,"_208":242,"_210":243},{"_186":239,"_24":207,"_208":240,"_210":214},{"_186":236,"_24":207,"_208":237,"_210":238},"fc2d92587766",[225],"Very briefly, you can't have tools that come back with success or failure.","eb03e88dbd0b",[],"a41910b828c5",[227],"and in order for long horizon agents to work, the tools have to be built for those use cases.","df8dd87193a8",[],"2d8f66dce2d4",[229],"And last but not least, and this is maybe a kind of catch-all category, so in order for, like I said, agents to interact with the outside world using tools,","b17b757adaf0",[258,259,260],[252,253],{"_186":256,"_24":198,"_199":257},{"_186":254,"_24":198,"_199":255},"f48b328","00:07:10.020","394093c","00:07:05.460",{"_186":266,"_24":207,"_208":267,"_210":268},{"_186":264,"_24":207,"_208":265,"_210":214},{"_186":261,"_24":207,"_208":262,"_210":263},"5dc259c16561",[254],"Just suffice to say that the model itself has to also take some of these experiences and affect its behavior from there.","2464553ff98a",[],"481cd7df48a9",[256],"The concept of learning is talked about a lot, and it is tied to memory.","87137fd1b59b",[278,279,280],[272,273],{"_186":276,"_24":198,"_199":277},{"_186":274,"_24":198,"_199":275},"c5a7398","00:07:00.259","b4fdf8d","00:06:55.460",{"_186":286,"_24":207,"_208":287,"_210":288},{"_186":284,"_24":207,"_208":285,"_210":214},{"_186":281,"_24":207,"_208":282,"_210":283},"45c146125494",[274],"And these have to be able to be resilient enough to work across long horizons.","cc906b91ced5",[],"b0c7e0ee6a90",[276],"So they need to solve these real-world problems using a computer or a sandbox.","4395c3d647da",[301,302,303,304,305],[292,293,294],{"_186":299,"_24":198,"_199":300},{"_186":297,"_24":198,"_199":298},{"_186":295,"_24":198,"_199":296},"0e59357","00:06:53.020","384dcaf","00:06:48.700","fc05f64","00:06:45.900",{"_186":316,"_24":207,"_208":317,"_210":318},{"_186":314,"_24":207,"_208":315,"_210":214},{"_186":311,"_24":207,"_208":312,"_210":313},{"_186":309,"_24":207,"_208":310,"_210":214},{"_186":306,"_24":207,"_208":307,"_210":308},"49b00cd9dfac",[295],"They interact with the world with a computer.","ba6f61fc5ed5",[],"be7aba518e30",[297],"So our agents today and our intelligence today live inside of computers.","b8359002e174",[],"a531292ad050",[299],"The third of these is kind of durable execution.","bfe01242f622",[331,332,333,334,335],[322,323,324],{"_186":329,"_24":198,"_199":330},{"_186":327,"_24":198,"_199":328},{"_186":325,"_24":198,"_199":326},"415b85a","00:06:40.820","8ad0b79","00:06:39.060","781ffba","00:06:34.420",{"_186":346,"_24":207,"_208":347,"_210":348},{"_186":344,"_24":207,"_208":345,"_210":214},{"_186":341,"_24":207,"_208":342,"_210":343},{"_186":339,"_24":207,"_208":340,"_210":214},{"_186":336,"_24":207,"_208":337,"_210":338},"bf0df3d0c08c",[325],"How do you make sure that short-term goals are aligned with long-term goals?","36813c18d2e8",[],"3a1369ba803b",[327],"How do you make sure the goals don't drift?","d65cd3a71250",[],"9e376789df1e",[329],"But how do you orchestrate goals across a large number of agents?","a22851a7b9e0",[355],[352],{"_186":353,"_24":198,"_199":354},"8010ea4","00:06:30.360",{"_186":356,"_24":207,"_208":357,"_210":358},"7dc330f1c61c",[353],"The second is goal orchestration. I touched on state decay.","633ff6400743",[371,372,373,374,375],[362,363,364],{"_186":369,"_24":198,"_199":370},{"_186":367,"_24":198,"_199":368},{"_186":365,"_24":198,"_199":366},"c79c7c5","00:06:23.840","dd58181","00:06:18.639","c72f221","00:06:16.639",{"_186":386,"_24":207,"_208":387,"_210":388},{"_186":384,"_24":207,"_208":385,"_210":214},{"_186":381,"_24":207,"_208":382,"_210":383},{"_186":379,"_24":207,"_208":380,"_210":214},{"_186":376,"_24":207,"_208":377,"_210":378},"04481ddbcdb6",[365],"Memory should allow the harness to affect future behavior based off of past experience.","f50d6dec4adc",[],"dfc354c22abf",[367],"A very simplistic view of memory is to think about it as the storage and recall, but memory is more than that.","946e748c2475",[],"36f0b8a9269d",[369],"The first of these is memory.","5d67c9c14bc4",[401,402,403,404,405],[392,393,394],{"_186":399,"_24":198,"_199":400},{"_186":397,"_24":198,"_199":398},{"_186":395,"_24":198,"_199":396},"50e28d4","00:06:11.039","ad5cd2f","00:06:02.520","1ca5d54","00:05:52.580",{"_186":416,"_24":207,"_208":417,"_210":418},{"_186":414,"_24":207,"_208":415,"_210":214},{"_186":411,"_24":207,"_208":412,"_210":413},{"_186":409,"_24":207,"_208":410,"_210":214},{"_186":406,"_24":207,"_208":407,"_210":408},"9a03fd4fb9f9",[395],"I'll very briefly touch on this, but you're gonna hear a lot more about these in the remainder of the day.","2145c9ac7b37",[],"4092d50a5d85",[397],"And again, a lot of these exist today, but they need to mature further before building long horizon agents becomes accessible to most engineers and organizations.","c6d09c7a83bb",[],"fd3154191601",[399],"And just to tee up some of the other kind of infrastructure pieces we care about, and I wrote about this in a blog post earlier in the summer, there's kind of five systems that need to mature.","b86feaa5532c",[423],[],"h2",{"_186":424,"_24":207,"_208":425,"_210":426},"6d74cf750771",[],"The Five Systems","08f1ef4dff2f",[436,437,438],[430,431],{"_186":434,"_24":198,"_199":435},{"_186":432,"_24":198,"_199":433},"255d3cc","00:05:43.460","1bf3dd5","00:05:35.080",{"_186":444,"_24":207,"_208":445,"_210":446},{"_186":442,"_24":207,"_208":443,"_210":214},{"_186":439,"_24":207,"_208":440,"_210":441},"44cfdc8ff0a8",[432],"is an open discussion of where each of these capabilities live, or if they live across these, how do they collaborate with each other?","f6337b81125c",[],"26586a8e1b9d",[434],"Now, those tools themselves might have models within themselves, they might be stochastic compute, stochastic programs, but there","0b59b175b3cc",[456,457,458],[450,451],{"_186":454,"_24":198,"_199":455},{"_186":452,"_24":198,"_199":453},"7ce0eb3","00:05:27.639","3ec1164","00:05:18.440",{"_186":464,"_24":207,"_208":465,"_210":466},{"_186":462,"_24":207,"_208":463,"_210":214},{"_186":459,"_24":207,"_208":460,"_210":461},"74977262cd7d",[452],"A typical agent, I mean I'm simplifying a lot here, is built around a model, has a harness, and has certain tools.","8dbbb7c22bda",[],"b3e6e90ed787",[454],"A major theme that we discuss, and I'm sure you're gonna hear a lot of in the remainder of the day, is where should some of these capabilities and infrastructure live?","32948014abcb",[470],[],{"_186":471,"_24":207,"_208":472,"_210":473},"f209446fc6c4",[],"Where Capabilities Should Live","73bd76bd520a",[486,487,488,489,490],[477,478,479],{"_186":484,"_24":198,"_199":485},{"_186":482,"_24":198,"_199":483},{"_186":480,"_24":198,"_199":481},"2b08c3f","00:05:09.080","340c788","00:04:59.940","194dcf0","00:04:50.340",{"_186":501,"_24":207,"_208":502,"_210":503},{"_186":499,"_24":207,"_208":500,"_210":214},{"_186":496,"_24":207,"_208":497,"_210":498},{"_186":494,"_24":207,"_208":495,"_210":214},{"_186":491,"_24":207,"_208":492,"_210":493},"b029febadd9d",[480],"And you need to do context engineering and you need tools to keep these things working across long horizons.","efab50d9a05e",[],"9bad787c7539",[482],"And so even across long horizons, even if you keep the long-term goals in the context, the agent over time starts drifting away from it.","23674100a243",[],"34943deb16cc",[484],"And if you're interested in reading more, the Meta Superintelligence team has a paper titled Remembering When It Matters, where they talk about this concept of state decay.","af5284ea2b7a",[516,517,518,519,520],[507,508,509],{"_186":514,"_24":198,"_199":515},{"_186":512,"_24":198,"_199":513},{"_186":510,"_24":198,"_199":511},"a08d646","00:04:47.620","771abf0","00:04:40.740","6adb3f2","00:04:33.740",{"_186":531,"_24":207,"_208":532,"_210":533},{"_186":529,"_24":207,"_208":530,"_210":214},{"_186":526,"_24":207,"_208":527,"_210":528},{"_186":524,"_24":207,"_208":525,"_210":214},{"_186":521,"_24":207,"_208":522,"_210":523},"73f981c3b6fd",[510],"We now know that approach on its own won't work.","624ef5fe8859",[],"266698a414c0",[512],"And people thought that the labs are gonna add context window and over time this is all gonna work and we're gonna have long horizon agents.","adba7ff9b275",[],"5869715148bb",[514],"And then back then, if some of you remember, the problem was seeing that we don't have enough context window.","4337b7f42b1b",[549,550,551,552,553,554,555],[537,538,539,540],{"_186":547,"_24":198,"_199":548},{"_186":545,"_24":198,"_199":546},{"_186":543,"_24":198,"_199":544},{"_186":541,"_24":198,"_199":542},"18909b9","00:04:32.380","af77473","00:04:27.720","ec120e3","00:04:23.880","16144e7","00:04:23.160",{"_186":571,"_24":207,"_208":572,"_210":573},{"_186":569,"_24":207,"_208":570,"_210":214},{"_186":566,"_24":207,"_208":567,"_210":568},{"_186":564,"_24":207,"_208":565,"_210":214},{"_186":561,"_24":207,"_208":562,"_210":563},{"_186":559,"_24":207,"_208":560,"_210":214},{"_186":556,"_24":207,"_208":557,"_210":558},"d1fee3653671",[541],"We're gonna put it in a loop.","f7c8861fe3d3",[],"98e579ec59f1",[543],"The first assumption was, okay, we're gonna take this model, we're gonna give it some tools to interact with the outside world.","51f81eb49a8f",[],"76ac8e9a93f4",[545],"I talked about how long horizon agents, kind of with AutoGPT.","e54e74c6b765",[],"ecec9b765b7f",[547],"Okay.","6a07ebd595f5",[583,584,585],[577,578],{"_186":581,"_24":198,"_199":582},{"_186":579,"_24":198,"_199":580},"255c467","00:04:12.140","36d26b5","00:04:06.620",{"_186":591,"_24":207,"_208":592,"_210":593},{"_186":589,"_24":207,"_208":590,"_210":214},{"_186":586,"_24":207,"_208":587,"_210":588},"edfb4dd11f92",[579],"horizon agents can work, but in order to make it accessible for organizations and engineers to do so, we still need some pieces of infrastructure to exist.","d9adcce8efbe",[],"89f44a138428",[581],"But now we, only a few years later, are here and we believe that long","3f77e06065f7",[603,604,605],[597,598],{"_186":601,"_24":198,"_199":602},{"_186":599,"_24":198,"_199":600},"e9a99ac","00:04:03.660","ec987e1","00:03:57.099",{"_186":611,"_24":207,"_208":612,"_210":613},{"_186":609,"_24":207,"_208":610,"_210":214},{"_186":606,"_24":207,"_208":607,"_210":608},"eb0dee837054",[599],"We needed models to do things like reasoning.","2cf5744aecbf",[],"d8ecb4ea4c85",[601],"But while the same idea still applies, it didn't work back then because we needed things like reflection.","692015eda2b8",[623,624,625],[617,618],{"_186":621,"_24":198,"_199":622},{"_186":619,"_24":198,"_199":620},"763ad26","00:03:55.000","b5dc6e8","00:03:46.040",{"_186":631,"_24":207,"_208":632,"_210":633},{"_186":629,"_24":207,"_208":630,"_210":214},{"_186":626,"_24":207,"_208":627,"_210":628},"23aa362035b1",[619],"And that's a whole different discussion.","3b09e04edfa6",[],"8744c118f2d0",[621],"And AutoGPT was engineers trying to build long horizon agents using this new intelligence that we had invented or discovered.","b308cd172fa7",[646,647,648,649,650],[637,638,639],{"_186":644,"_24":198,"_199":645},{"_186":642,"_24":198,"_199":643},{"_186":640,"_24":198,"_199":641},"1ea4bc0","00:03:43.000","a83d55d","00:03:32.840","db9307c","00:03:25.700",{"_186":661,"_24":207,"_208":662,"_210":663},{"_186":659,"_24":207,"_208":660,"_210":214},{"_186":656,"_24":207,"_208":657,"_210":658},{"_186":654,"_24":207,"_208":655,"_210":214},{"_186":651,"_24":207,"_208":652,"_210":653},"461b30e9867b",[640],"people tried to do AutoGPT.","03e23cd49628",[],"ab86e50ec4e0",[642],"But from November 2022, when ChatGPT really came out and got a lot of people's attention, only a few months later, like March 2023,","e300dbf32eea",[],"d0eedb36ce5e",[644],"But what I'm trying to show to you here is that people were trying transformers, text in, text out.","55c37e5a7335",[676,677,678,679,680],[667,668,669],{"_186":674,"_24":198,"_199":675},{"_186":672,"_24":198,"_199":673},{"_186":670,"_24":198,"_199":671},"d1264e0","00:03:18.820","4eee369","00:03:13.600","367dea6","00:03:04.700",{"_186":691,"_24":207,"_208":692,"_210":693},{"_186":689,"_24":207,"_208":690,"_210":214},{"_186":686,"_24":207,"_208":687,"_210":688},{"_186":684,"_24":207,"_208":685,"_210":214},{"_186":681,"_24":207,"_208":682,"_210":683},"e29aa3966a89",[670],"I'm not gonna bore you here with a history of LLMs and AI, which most of you know.","1014bb6cdb81",[],"13c036ad5bf8",[672],"before everyone, all engineers and all organizations, can go and deploy these long horizon agents.","756d74e4934e",[],"3904c32759af",[674],"But I want to talk a little bit about how we got here and tee up some ideas of what we think needs to happen","9345e92203b7",[697],[],{"_186":698,"_24":207,"_208":699,"_210":700},"72e30c3a7827",[],"How We Got Here","1d1c5cb58cf7",[707],[704],{"_186":705,"_24":198,"_199":706},"11c0ccf","00:02:54.459",{"_186":708,"_24":207,"_208":709,"_210":710},"00ee6bf2c2c3",[705],"Now, we can discuss whether math is verifiable, not everything is like math, but for me, this shows that long horizon agents can now work.","26c2d0e16dc0",[720,721,722],[714,715],{"_186":718,"_24":198,"_199":719},{"_186":716,"_24":198,"_199":717},"00f7001","00:02:48.220","9abfb4a","00:02:43.280",{"_186":728,"_24":207,"_208":729,"_210":730},{"_186":726,"_24":207,"_208":727,"_210":214},{"_186":723,"_24":207,"_208":724,"_210":725},"3d6e39860e89",[716],"they would have to work eight hours a day for 4,000 straight years to solve this problem.","27796027fa16",[],"ee5b74deae0d",[718],"If someone like Grigori Perelman had wanted to go and try to solve this in this way,","58095447d0df",[740,741,742],[734,735],{"_186":738,"_24":198,"_199":739},{"_186":736,"_24":198,"_199":737},"cad579b","00:02:37.200","4031532","00:02:26.840",{"_186":748,"_24":207,"_208":749,"_210":750},{"_186":746,"_24":207,"_208":747,"_210":214},{"_186":743,"_24":207,"_208":744,"_210":745},"44c5181161cb",[736],"generating 130 billion tokens to solve this problem.","ee772978e4a5",[],"57d3a7b7315d",[738],"So OpenAI took one of their next generation models and unleashed 10,000 agents that worked in tandem for 88 hours,","a45d888806f4",[766,767,768,769,770,771,772],[754,755,756,757],{"_186":764,"_24":198,"_199":765},{"_186":762,"_24":198,"_199":763},{"_186":760,"_24":198,"_199":761},{"_186":758,"_24":198,"_199":759},"074e0c8","00:02:22.220","56deffb","00:02:10.740","71cb158","00:02:08.619","9ad52e4","00:01:57.340",{"_186":788,"_24":207,"_208":789,"_210":790},{"_186":786,"_24":207,"_208":787,"_210":214},{"_186":783,"_24":207,"_208":784,"_210":785},{"_186":781,"_24":207,"_208":782,"_210":214},{"_186":778,"_24":207,"_208":779,"_210":780},{"_186":776,"_24":207,"_208":777,"_210":214},{"_186":773,"_24":207,"_208":774,"_210":775},"a0a507d6e086",[758],"the amount of cognitive power that was thrown at this problem.","1c0f4cfe633a",[],"5bf5f85d4396",[760],"to understand why this clearly shows how long horizon agents can work across very long horizons, I'm gonna share a few numbers with you about","88740a008092",[],"1e5185837b6b",[762],"Now,","c08f7b4adf3d",[],"c201df309040",[764],"And what happened then was another one of these seven problems, known as the Navier-Stokes equations, a solution was proposed, which is also verified, by OpenAI.","2bddb6321e77",[800,801,802],[794,795],{"_186":798,"_24":198,"_199":799},{"_186":796,"_24":198,"_199":797},"1bb45e4","00:01:49.659","1c98bc5","00:01:45.259",{"_186":808,"_24":207,"_208":809,"_210":810},{"_186":806,"_24":207,"_208":807,"_210":214},{"_186":803,"_24":207,"_208":804,"_210":805},"7145795a1d40",[796],"And of the other six, no one had been able to solve any of them, but that was until two weeks ago.","4d88463c80c1",[],"0542e4bc4948",[798],"No one heard of him until he came out after eight years with the proof of that problem.","d9cc22dda207",[820,821,822],[814,815],{"_186":818,"_24":198,"_199":819},{"_186":816,"_24":198,"_199":817},"61c77c9","00:01:34.939","22a2341","00:01:26.600",{"_186":828,"_24":207,"_208":829,"_210":830},{"_186":826,"_24":207,"_208":827,"_210":214},{"_186":823,"_24":207,"_208":824,"_210":825},"79a87162ab69",[816],"Only one of them was proven in the past twenty-six years and it was done by this gentleman called Grigori Perelman who locked himself away in a room effectively for eight years.","2aa8d0582c5e",[],"a2fbde00fd56",[818],"And the point is that of those seven problems over the past 26 years, a lot of humans have tried to solve these or prove these.","fe6dcc2a82d2",[843,844,845,846,847],[834,835,836],{"_186":841,"_24":198,"_199":842},{"_186":839,"_24":198,"_199":840},{"_186":837,"_24":198,"_199":838},"e27921b","00:01:22.540","501bdf2","00:01:13.380","26368ed","00:01:10.580",{"_186":858,"_24":207,"_208":859,"_210":860},{"_186":856,"_24":207,"_208":857,"_210":214},{"_186":853,"_24":207,"_208":854,"_210":855},{"_186":851,"_24":207,"_208":852,"_210":214},{"_186":848,"_24":207,"_208":849,"_210":850},"da210c59f008",[837],"And so why am I talking about all this? It's to make a point.","bdf79a3e0591",[],"6e2dc18eb577",[839],"And for me, this over the past 26 years represented really the epitome of what human cognition could try to solve, attempt to solve.","c11d4f4c8432",[],"2f80d5334105",[841],"It tells us about the distribution of the primes.","6302df2e60d7",[870,871,872],[864,865],{"_186":868,"_24":198,"_199":869},{"_186":866,"_24":198,"_199":867},"e719f77","00:01:04.000","4b70d72","00:00:53.600",{"_186":878,"_24":207,"_208":879,"_210":880},{"_186":876,"_24":207,"_208":877,"_210":214},{"_186":873,"_24":207,"_208":874,"_210":875},"34a6d9878849",[866],"And these problems have been around for a number of years. This is the one I really like the most, for those few years that I was obsessed with it.","9e0c79cdcb2a",[],"1b65e9cb52a6",[868],"I was a college student 20 years ago and the Clay Institute of Mathematics basically picked seven math problems as being the most important open problems in mathematics at the time.","aa5f3492319a",[890,891,892],[884,885],{"_186":888,"_24":198,"_199":889},{"_186":886,"_24":198,"_199":887},"838777b","00:00:48.500","84e903b","00:00:42.579",{"_186":898,"_24":207,"_208":899,"_210":900},{"_186":896,"_24":207,"_208":897,"_210":214},{"_186":893,"_24":207,"_208":894,"_210":895},"5b920c31aceb",[886],"This is a representation of what's known as the Riemann zeta function.","95f15ca2cf35",[],"6ebe25c390a9",[888],"Pretty much what I expected. So the graph here, I like to think of this as my spirit symbol.","edc7719328fc",[910,911,912],[904,905],{"_186":908,"_24":198,"_199":909},{"_186":906,"_24":198,"_199":907},"0a517ae","00:00:40.660","d032dfb","00:00:32.660",{"_186":918,"_24":207,"_208":919,"_210":920},{"_186":916,"_24":207,"_208":917,"_210":214},{"_186":913,"_24":207,"_208":914,"_210":915},"6b5573ba58d2",[906],"One, two, three people. All right.","3cfa96babbbd",[],"7673e8081e66",[908],"So, quick show of hands. Does anyone know what this graph here represents?","2c4fdd173fbf",[930,931,932],[924,925],{"_186":928,"_24":198,"_199":929},{"_186":926,"_24":198,"_199":927},"d84bca4","00:00:27.460","3004b1c","00:00:19.000",{"_186":938,"_24":207,"_208":939,"_210":940},{"_186":936,"_24":207,"_208":937,"_210":214},{"_186":933,"_24":207,"_208":934,"_210":935},"420bb9111670",[926],"And the only way humans can prove anything to each other, that's using math.","4b520828029e",[],"4e8ac6b28730",[928],"I believe that long horizon agents can now work, and the next few slides, I'm gonna prove that to you, or at least attempt to prove that to you.","d063787f5097",[950,951,952],[944,945],{"_186":948,"_24":198,"_199":949},{"_186":946,"_24":198,"_199":947},"3033d42","00:00:15.240","32eb6b8","00:00:04.400",{"_186":958,"_24":207,"_208":959,"_210":960},{"_186":956,"_24":207,"_208":957,"_210":214},{"_186":953,"_24":207,"_208":954,"_210":955},"3fd27eab43fa",[946],"And the reason for me is pretty clear.","93d2b875981b",[],"ceb0a1dcc8a8",[948],"Now before we start with our first panel, I'm gonna talk a little bit about why we picked the Long Horizon Stack as the theme for this DevGuild.","f215dfe363d6",[964],[],{"_186":965,"_24":207,"_208":966,"_210":967},"f12d4354e92b",[],"A Proof From Mathematics",{"_18":1019,"_20":1020,"_22":1021,"_24":978,"_26":1019,"_979":1022,"_981":1023,"_14":1024},{"_18":986,"_20":1014,"_22":988,"_24":978,"_26":989,"_979":1015,"_981":1016,"_14":1017},{"_18":1006,"_20":1007,"_22":1008,"_24":978,"_26":1009,"_979":1010,"_981":1011,"_14":1012},{"_18":986,"_20":1001,"_22":988,"_24":978,"_26":989,"_979":1002,"_981":1003,"_14":1004},{"_18":994,"_20":995,"_22":996,"_24":978,"_26":994,"_979":997,"_981":998,"_14":999},{"_18":986,"_20":987,"_22":988,"_24":978,"_26":989,"_979":990,"_981":991,"_14":992},{"_18":975,"_20":976,"_22":977,"_24":978,"_26":975,"_979":980,"_981":982,"_14":983},"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,"_984":985},"current","ai","2026-10-23T23:10:26Z","8d2828f0-70fa-49f2-a7cf-4b648e941299","mG7vmCjaB1vEw3gqUIqrjf","2026-10-23T23:11:12Z","Coordinating many agents, goals, and roles.","Agent Orchestration",{"_24":14,"_984":993},"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,"_984":1000},"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,"_984":1005},"long-horizon-agents","2025-12-09T17:45:33Z","2a3f44e0-7a12-4730-a2e2-695a100b75be","6ACXEUwYcGP2Q5rQl5ATFI","2025-12-09T17:47:31Z","Context engineering is the process of designing the information payload entered into a prompt to improve LLMs' chances of efficiently returning an accurate response.","Context Engineering",{"_24":14,"_984":1013},"context-engineering","1fc6fbde-4346-46ff-84e3-43a620e9a3f4","How agents store, recall, forget, and learn from past work.","Agent Memory",{"_24":14,"_984":1018},"agent-memory","2024-09-05T17:31:35Z","017fa9be-32e2-46ee-bddf-f13275fbdf00","aeN0todGvIjkWyTywFzV6F","Pertaining to the technology underpinning AI and machine learning operations.","AI Infrastructure",{"_24":14,"_984":1025},"ai-infrastructure",{"_186":1027,"_24":1028,"_1029":-5,"_1030":1031},"spk0amirz","speaker","organization","person",{"_18":1032,"_20":1033,"_22":1034,"_1035":1036,"_24":1030,"_26":1037,"_1038":1039,"_1040":31,"_44":1041,"_981":1042,"_1029":1043,"_1044":31,"_1045":1046,"_1047":1048,"_68":1049,"_14":1050,"_1051":1052,"_61":1053,"_1054":31},"2025-07-23T22:53:13Z","7fbf3d51-584f-4e07-a1af-1b3d207c043a","FShviDYV4x5sjN0LtjCvP1","_system",{"_1182":1183},"2026-10-14T15:31:34Z","bio",[1174],"generate",[1128,1129],"Amir Zohrenejad",{"_18":1121,"_20":1122,"_22":1123,"_24":1029,"_26":1121,"_1040":12,"_981":1124,"_1125":12,"_68":1126},"partner","photo",{"_24":1094,"_1095":1096,"_74":1097,"_1098":1099,"_1100":1101},"position","General Partner",{"_72":1086,"_24":68,"_979":1087,"_1088":12,"_1089":12,"_63":1090},{"_24":14,"_984":1085},"socials",[1075,1076],[1055,1056,1057,1058,1059,1060],"teamMember",{"_186":1073,"_1062":1074,"_24":1064},{"_186":1071,"_1062":1072,"_24":1064},{"_186":1069,"_1062":1070,"_24":1064},{"_186":1067,"_1062":1068,"_24":1064},{"_186":1065,"_1062":1066,"_24":1064},{"_186":1061,"_1062":1063,"_24":1064},"d5cbb9dc34e9","_ref","kFpYkfzWCiYI8xC60oUnHC","reference","ae520a69275a","kFpYkfzWCiYI8xC60oUXoK","e3259c43bc5b","kFpYkfzWCiYI8xC60oUh0y","6c8821665ba3","35d87736-5801-4207-b96c-abf2f3cabf25","ee3f931d46f1","recL3C74SyXHVzCA2","2b3548c3ee0c","recqBryoYgSAx9YcB",{"_186":1082,"_24":1078,"_1079":1083,"_981":1084},{"_186":1077,"_24":1078,"_1079":1080,"_981":1081},"95f8c0026fdc","socialMedia","link","https://x.com/amirzo","twitter","beb63c6a9a03","https://www.linkedin.com/in/amir-zohrenejad/","linkedin","amir-zohrenejad",{"_24":1091,"_74":1092},"Amir Zohrenejad is a part-time partner at Heavybit and a serial entrepreneur who co-founded two technical startups: Dataherald and CloudFish.","nofollow","noindex","Amir Zohrenejad, Part-Time Partner at Heavybit","image",{"_1062":1093,"_24":1064},"image-a5bc5658d2ee2a57a138b1c416349e03363fcab0-1200x630-png","mainImage","alt","Amir Zohrenejad's Headshot",{"_1062":1120,"_24":1064},"crop",{"_24":1111,"_1112":1113,"_1114":1115,"_1116":1117,"_1118":1119},"hotspot",{"_24":1102,"_1103":1104,"_1105":1106,"_1107":1108,"_1109":1110},"sanity.imageHotspot","height",0.5134356192252267,"width",0.5074308563412813,"x",0.464549697141559,"y",0.30786428753072814,"sanity.imageCrop","bottom",0.4354179028566585,"left",0.21083426897091836,"right",0.28173487468780034,"top",0.051146477918114774,"image-a516037bcbe74034f1006d758dac5b252ea2d957-4800x6000-jpg","2022-07-21T16:35:12Z","fb878f3a-5a88-4ee9-a3d3-927c92b46a31","aeN0todGvIjkWyTywFzauG","Heavybit","portfolioCompany",{"_24":68,"_979":1127,"_1088":12,"_1089":12,"_63":1124},"The leading investor in developer-first startups",{"_186":1146,"_24":1147,"_210":1148,"_63":1149},{"_186":1130,"_24":1131,"_1132":1133},"80bbd63367fc","testimonialsCarousel","testimonials",[1134,1135,1136,1137],{"_186":1144,"_1062":1145,"_24":1064},{"_186":1142,"_1062":1143,"_24":1064},{"_186":1140,"_1062":1141,"_24":1064},{"_186":1138,"_1062":1139,"_24":1064},"37820e088baa","afd3991e-9c82-41ec-9eec-4fb8cdd1430a","44a0a53f497d","968941aa-1e4a-4c03-975a-9b9c5b7373c0","aa2fdae7d637","138c4b92-8c01-44cc-a90c-d42778234014","b9a9fa22da55","e10176d1-b3de-4ff9-80af-c8b757de2be4","42be7f40c63d","textBlock",[1150,1151,1152],"About Amir Zohrenejad",{"_186":1167,"_24":188,"_189":1168,"_191":1169,"_193":221},{"_186":1160,"_24":188,"_189":1161,"_191":1162,"_193":221},{"_186":1153,"_24":188,"_189":1154,"_191":1155,"_193":221},"205821c1d926",[1156],[],{"_186":1157,"_24":207,"_208":1158,"_210":1159},"c9037693b85a",[],"Amir specializes in helping early-stage startups recruit top-tier ML researchers and build applied AI teams that can ship production-ready systems. He brings deep experience in natural language interfaces, database infrastructure, and agent evaluation methodologies. On the go-to-market side, he advises founders on developer-focused strategies, from open source adoption to navigating long enterprise sales cycles with technical buyers.","58c9d15086f7",[1163],[],{"_186":1164,"_24":207,"_208":1165,"_210":1166},"d76c6c73c4c4",[],"Before joining Heavybit, Amir was the co-founder and served as CTO of Dataherald (YC W21), an open source natural-language-to-SQL library. There, he recruited the researchers and implemented emerging applied AI methodologies that led to the company’s state-of-the-art solution.","e47f8a1e9757",[1170],[],{"_186":1171,"_24":207,"_208":1172,"_210":1173},"0aa7378490c7",[],"Amir Zohrenejad is a General Partner at Heavybit. He began his journey in software development at UC Berkeley, where he graduated with degrees in both Applied Mathematics and Electrical Engineering and Computer Science. After graduating, he joined Microsoft as a technical program manager on the Windows Wireless team.",{"_186":1175,"_24":188,"_189":1176,"_191":1177,"_193":221},"ff74c895e24a",[1178],[],{"_186":1179,"_24":207,"_208":1180,"_210":1181},"7dd27ef0d731",[],"Amir Zohrenejad is a General Partner at Heavybit and a serial entrepreneur who co-founded two technical startups: Dataherald and CloudFish. A lifelong enthusiast of software and startups, Amir is particularly passionate about relational databases, applied artificial intelligence, and small language models. He enjoys helping ambitious early-stage founders navigate go-to-market challenges.","base",{"_100":1033,"_1184":1185},"rev","yFbD6tcnJso7qIzjSkcMqY",{"_1062":1187,"_24":1064},"image-fdb3c126c6efdae6a50918f5ed42892ce8186474-1920x1080-jpg",{"_186":1189,"_24":188,"_189":1190,"_191":1191,"_193":221},"22a0c7f4bcf6",[1192],[],{"_186":1193,"_24":207,"_208":1194,"_210":1195},"9d016f47a35b",[],"

The Long-Horizon Stack\n
- A Proof From Mathematics
- How We Got Here
- Where Capabilities Should Live
- The Five Systems
\n

",{"_186":1203,"_1062":1204,"_24":1064},{"_186":1201,"_1062":1202,"_24":1064},{"_186":1199,"_1062":1200,"_24":1064},"MYNusMFiOrRV","84a19c24-d9a7-4f56-b85b-400f11100e63","IhcMHctGGhFq","88a8c3a5-ecd1-4f4e-84ac-cb170d8849f4","h850fcNFQWUi","c491146a-c0b5-47e7-b78b-acecf8893ef0","items",[3668,3669,3670],"recent",[1209,1210,1211],{"_1212":3050,"_96":3051,"_1091":-5,"_1216":3052,"_51":3053,"_14":3054,"_1220":1221,"_63":3055,"_1223":1224},{"_1212":1830,"_96":1831,"_1091":-5,"_1216":1832,"_51":1833,"_14":1834,"_1220":1221,"_63":1835,"_1223":1224},{"_1212":1213,"_96":1214,"_1091":1215,"_1216":1217,"_51":1218,"_14":1219,"_1220":1221,"_63":1222,"_1223":1224},"content",[1225,1226,1227,1228,1229,1230,1231,1232,1233,1234,1235,1236,1237,1238,1239,1240,1241,1242,1243,1244,1245,1246,1247,1248,1249,1250,1251,1252,1253,1254,1255,1256,1257,1258,1259,1260,1261,1262,1263,1264,1265,1266,1267,1268,1269],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",{"_186":1823,"_24":188,"_189":1824,"_191":1825,"_193":422},{"_186":1794,"_24":188,"_189":1795,"_191":1796,"_193":221},{"_186":1765,"_24":188,"_189":1766,"_191":1767,"_193":221},{"_186":1762,"_24":1091,"_74":1763},{"_186":1749,"_24":188,"_189":1750,"_191":1751,"_193":221},{"_186":1742,"_24":188,"_189":1743,"_191":1744,"_193":422},{"_186":1724,"_24":188,"_189":1725,"_191":1726,"_193":221},{"_186":1706,"_24":188,"_189":1707,"_191":1708,"_193":221},{"_186":1699,"_24":188,"_189":1700,"_191":1701,"_193":221},{"_186":1692,"_24":188,"_189":1693,"_191":1694,"_193":221},{"_186":1685,"_24":188,"_189":1686,"_191":1687,"_193":422},{"_186":1656,"_24":188,"_189":1657,"_191":1658,"_193":221},{"_186":1649,"_24":188,"_189":1650,"_191":1651,"_193":221},{"_186":1642,"_24":188,"_189":1643,"_191":1644,"_193":221},{"_186":1635,"_24":188,"_189":1636,"_191":1637,"_193":422},{"_186":1628,"_24":188,"_189":1629,"_191":1630,"_193":221},{"_186":1610,"_24":188,"_189":1611,"_191":1612,"_193":221},{"_186":1592,"_24":188,"_189":1593,"_191":1594,"_193":221},{"_186":1585,"_24":188,"_189":1586,"_191":1587,"_193":422},{"_186":1578,"_24":188,"_189":1579,"_191":1580,"_193":221},{"_186":1552,"_24":188,"_189":1553,"_191":1554,"_193":221},{"_186":1534,"_24":188,"_189":1535,"_191":1536,"_193":221},{"_186":1519,"_24":188,"_189":1520,"_191":1521,"_193":221},{"_186":1512,"_24":188,"_189":1513,"_191":1514,"_193":221},{"_186":1505,"_24":188,"_189":1506,"_191":1507,"_193":221},{"_186":1498,"_24":188,"_189":1499,"_191":1500,"_193":422},{"_186":1491,"_24":188,"_189":1492,"_191":1493,"_193":221},{"_186":1484,"_24":188,"_189":1485,"_191":1486,"_193":221},{"_186":1477,"_24":188,"_189":1478,"_191":1479,"_193":221},{"_186":1448,"_24":188,"_189":1449,"_191":1450,"_193":221},{"_186":1441,"_24":188,"_189":1442,"_191":1443,"_193":422},{"_186":1434,"_24":188,"_189":1435,"_191":1436,"_193":221},{"_186":1427,"_24":188,"_189":1428,"_191":1429,"_193":221},{"_186":1420,"_24":188,"_189":1421,"_191":1422,"_193":221},{"_186":1413,"_24":188,"_189":1414,"_191":1415,"_193":221},{"_186":1406,"_24":188,"_189":1407,"_191":1408,"_193":422},{"_186":1399,"_24":188,"_189":1400,"_191":1401,"_193":221},{"_186":1359,"_24":188,"_189":1360,"_191":1361,"_193":221},{"_186":1329,"_24":188,"_189":1330,"_191":1331,"_193":221},{"_186":1322,"_24":188,"_189":1323,"_191":1324,"_193":422},{"_186":1315,"_24":188,"_189":1316,"_191":1317,"_193":221},{"_186":1300,"_24":188,"_189":1301,"_191":1302,"_193":221},{"_186":1293,"_24":188,"_189":1294,"_191":1295,"_193":221},{"_186":1277,"_24":188,"_189":1278,"_191":1279,"_193":221},{"_186":1270,"_24":188,"_189":1271,"_191":1272,"_193":221},"8eaf5dda166e",[1273],[],{"_186":1274,"_24":207,"_208":1275,"_210":1276},"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",[1280,1281,1282],[],{"_186":1290,"_24":207,"_208":1291,"_210":1292},{"_186":1286,"_24":207,"_208":1287,"_210":1288},{"_186":1283,"_24":207,"_208":1284,"_210":1285},"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",[1289],"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",[1296],[],{"_186":1297,"_24":207,"_208":1298,"_210":1299},"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",[1303,1304,1305],[],{"_186":1312,"_24":207,"_208":1313,"_210":1314},{"_186":1309,"_24":207,"_208":1310,"_210":1311},{"_186":1306,"_24":207,"_208":1307,"_210":1308},"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",[1289],"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",[1318],[],{"_186":1319,"_24":207,"_208":1320,"_210":1321},"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",[1325],[],{"_186":1326,"_24":207,"_208":1327,"_210":1328},"2586e6d9e8b9",[],"The Future of Codebases: Built and Managed by AI?","6869a525750a",[1339,1340,1341,1342,1343],[1332,1333],{"_186":1337,"_24":1079,"_1335":1338},{"_186":1334,"_24":1079,"_1335":1336},"41760ee29615","href","https://www.youtube.com/watch?v=O2jUTq6nCEY&t=2s","48182a105088","https://blog.dailydoseofds.com/",{"_186":1356,"_24":207,"_208":1357,"_210":1358},{"_186":1353,"_24":207,"_208":1354,"_210":1355},{"_186":1350,"_24":207,"_208":1351,"_210":1352},{"_186":1347,"_24":207,"_208":1348,"_210":1349},{"_186":1344,"_24":207,"_208":1345,"_210":1346},"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",[1334],"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",[1337],"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",[1371,1372,1373,1374,1375,1376,1377],[1362,1363,1364],{"_186":1369,"_24":1079,"_1335":1370},{"_186":1367,"_24":1079,"_1335":1368},{"_186":1365,"_24":1079,"_1335":1366},"2f238ed440f7","https://zed.dev/","3f971d287b17","https://neovim.io/","f040745fb71d","https://github.com/vitali87/croft",{"_186":1396,"_24":207,"_208":1397,"_210":1398},{"_186":1393,"_24":207,"_208":1394,"_210":1395},{"_186":1390,"_24":207,"_208":1391,"_210":1392},{"_186":1387,"_24":207,"_208":1388,"_210":1389},{"_186":1384,"_24":207,"_208":1385,"_210":1386},{"_186":1381,"_24":207,"_208":1382,"_210":1383},{"_186":1378,"_24":207,"_208":1379,"_210":1380},"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",[1365],"Zed","e76bcfb5a8f3",[],", ","94d7ac6fa32a",[1367],"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",[1369],"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",[1402],[],{"_186":1403,"_24":207,"_208":1404,"_210":1405},"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",[1409],[],{"_186":1410,"_24":207,"_208":1411,"_210":1412},"76872200c049",[],"Learnings from Building","526da0bc67f6",[1416],[],{"_186":1417,"_24":207,"_208":1418,"_210":1419},"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",[1423],[],{"_186":1424,"_24":207,"_208":1425,"_210":1426},"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",[1430],[],{"_186":1431,"_24":207,"_208":1432,"_210":1433},"eb00873aac4a",[],"Avagyan notes that those companies may be hitting a stumbling block. “I think one problem companies may be having is accuracy.”","7cbaec07cf9b",[1437],[],{"_186":1438,"_24":207,"_208":1439,"_210":1440},"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",[1444],[],{"_186":1445,"_24":207,"_208":1446,"_210":1447},"1cd77c6dd0f6",[],"Why Aren’t Enterprises Already Graphing Their Codebases?","b003dd295bbf",[1457,1458,1459,1460,1461],[1451,1452],{"_186":1455,"_24":1079,"_1335":1456},{"_186":1453,"_24":1079,"_1335":1454},"eb5b74dd23fd","https://flutter.dev/","fc884ca87ef9","https://dart.dev/overview",{"_186":1474,"_24":207,"_208":1475,"_210":1476},{"_186":1471,"_24":207,"_208":1472,"_210":1473},{"_186":1468,"_24":207,"_208":1469,"_210":1470},{"_186":1465,"_24":207,"_208":1466,"_210":1467},{"_186":1462,"_24":207,"_208":1463,"_210":1464},"b630193673be",[]," widget structure.”","ab660c7a1ffd",[1453],"Flutter","cee0616ee44a",[]," is already supported and includes ","fc85ec3d18b6",[1455],"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",[1480],[],{"_186":1481,"_24":207,"_208":1482,"_210":1483},"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",[1487],[],{"_186":1488,"_24":207,"_208":1489,"_210":1490},"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",[1494],[],{"_186":1495,"_24":207,"_208":1496,"_210":1497},"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",[1501],[],{"_186":1502,"_24":207,"_208":1503,"_210":1504},"0743e3b412d6",[],"Real-World Use Cases for Graphing a Codebase","6a4ddfd50139",[1508],[],{"_186":1509,"_24":207,"_208":1510,"_210":1511},"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",[1515],[],{"_186":1516,"_24":207,"_208":1517,"_210":1518},"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",[1522,1523,1524],[],{"_186":1531,"_24":207,"_208":1532,"_210":1533},{"_186":1528,"_24":207,"_208":1529,"_210":1530},{"_186":1525,"_24":207,"_208":1526,"_210":1527},"e051511f75df",[]," say that I’ve observed the project to be very fast, and it seems more fault-tolerant in a way.”","36ec44554375",[1289],"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",[1540,1541,1542],[1537],{"_186":1538,"_24":1079,"_1335":1539},"c1ed526f5769","https://memgraph.com/",{"_186":1549,"_24":207,"_208":1550,"_210":1551},{"_186":1546,"_24":207,"_208":1547,"_210":1548},{"_186":1543,"_24":207,"_208":1544,"_210":1545},"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",[1538],"Memgraph","c9fdbc40e4d6",[],"“Otherwise, I don't think storage is a problem. The project uses ","d8236069c0ba",[1558,1559,1560,1561,1562],[1555],{"_186":1556,"_24":1079,"_1335":1557},"ec14de8b6db9","https://en.wikipedia.org/wiki/Dead-code_elimination",{"_186":1575,"_24":207,"_208":1576,"_210":1577},{"_186":1572,"_24":207,"_208":1573,"_210":1574},{"_186":1569,"_24":207,"_208":1570,"_210":1571},{"_186":1566,"_24":207,"_208":1567,"_210":1568},{"_186":1563,"_24":207,"_208":1564,"_210":1565},"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",[1289],"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",[1556],"dead code detection","37e4c8419854",[],"“I would say that a practical use of code-graph-rag is ","2e27e83e6ea0",[1581],[],{"_186":1582,"_24":207,"_208":1583,"_210":1584},"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",[1588],[],{"_186":1589,"_24":207,"_208":1590,"_210":1591},"8bfe3199c316",[],"Bottlenecks and Resource Management","9b53b53a2ef7",[1598,1599,1600],[1595],{"_186":1596,"_24":1079,"_1335":1597},"cb676615c060","https://simonwillison.net/guides/agentic-engineering-patterns/red-green-tdd/",{"_186":1607,"_24":207,"_208":1608,"_210":1609},{"_186":1604,"_24":207,"_208":1605,"_210":1606},{"_186":1601,"_24":207,"_208":1602,"_210":1603},"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",[1596],"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",[1616,1617,1618],[1613],{"_186":1614,"_24":1079,"_1335":1615},"1b1e54e0707e","https://en.wikipedia.org/wiki/Polymorphism_(computer_science)",{"_186":1625,"_24":207,"_208":1626,"_210":1627},{"_186":1622,"_24":207,"_208":1623,"_210":1624},{"_186":1619,"_24":207,"_208":1620,"_210":1621},"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",[1614],"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",[1631],[],{"_186":1632,"_24":207,"_208":1633,"_210":1634},"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",[1638],[],{"_186":1639,"_24":207,"_208":1640,"_210":1641},"1641bec46e7a",[],"Making Code Graphs Deterministic with AST","a36ee5a77fd2",[1645],[],{"_186":1646,"_24":207,"_208":1647,"_210":1648},"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",[1652],[],{"_186":1653,"_24":207,"_208":1654,"_210":1655},"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",[1665,1666,1667,1668,1669],[1659,1660],{"_186":1663,"_24":1079,"_1335":1664},{"_186":1661,"_24":1079,"_1335":1662},"249727404c78","https://en.wikipedia.org/wiki/Abstract_syntax_tree","d7c41e85c164","https://tree-sitter.github.io/tree-sitter/",{"_186":1682,"_24":207,"_208":1683,"_210":1684},{"_186":1679,"_24":207,"_208":1680,"_210":1681},{"_186":1676,"_24":207,"_208":1677,"_210":1678},{"_186":1673,"_24":207,"_208":1674,"_210":1675},{"_186":1670,"_24":207,"_208":1671,"_210":1672},"916d54525867",[]," and gives you all the syntactic relationships, it's not still enough because there’s so much variety between languages and frameworks.”","ecb02d6d29a7",[1661],"abstract syntax tree","2d414b50c68b",[],", which parses the ","c4f79c67c17a",[1663],"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",[1688],[],{"_186":1689,"_24":207,"_208":1690,"_210":1691},"a7a96359e940",[],"Accounting for Variation Across Languages, Frameworks","55289533b05a",[1695],[],{"_186":1696,"_24":207,"_208":1697,"_210":1698},"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",[1702],[],{"_186":1703,"_24":207,"_208":1704,"_210":1705},"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",[1712,1713,1714],[1709],{"_186":1710,"_24":1079,"_1335":1711},"40d5a9d1fab0","https://en.wikipedia.org/wiki/Cypher_(query_language)",{"_186":1721,"_24":207,"_208":1722,"_210":1723},{"_186":1718,"_24":207,"_208":1719,"_210":1720},{"_186":1715,"_24":207,"_208":1716,"_210":1717},"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",[1710],"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",[1730,1731,1732],[1727],{"_186":1728,"_24":1079,"_1335":1729},"67f4fbd2fe9a","https://en.wikipedia.org/wiki/Grep",{"_186":1739,"_24":207,"_208":1740,"_210":1741},{"_186":1736,"_24":207,"_208":1737,"_210":1738},{"_186":1733,"_24":207,"_208":1734,"_210":1735},"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",[1728],"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",[1745],[],{"_186":1746,"_24":207,"_208":1747,"_210":1748},"4cb4e6406d24",[],"Viewing Code as a Retrievable Graph","b80310699127",[1755,1756],[1752],{"_186":1753,"_24":1079,"_1335":1754},"5b72c578e5d1","https://www.youtube.com/watch?v=O2jUTq6nCEY",{"_186":1759,"_24":207,"_208":1760,"_210":1761},{"_186":1757,"_24":207,"_208":1758,"_210":1548},"5ef047d4ccf9",[1753,1289],"a0931bc06a15",[1289],"Vitali Avagyan discusses using GraphRAG to parse repos with Memgraph. Image courtesy ","c9d37b6922a5",{"_1062":1764,"_24":1064},"image-bc24d7b1e62740bb7448a1a80a9dfd59940c68d6-3840x1100-jpg","9b432548c002",[1774,1775,1776,1777,1778],[1768,1769],{"_186":1772,"_24":1079,"_1335":1773},{"_186":1770,"_24":1079,"_1335":1771},"b6ce27c00ce8","https://github.com/vitali87/code-graph-rag","a50030cc5b04","https://www.linkedin.com/in/vitali-avagyan-a1566234/",{"_186":1791,"_24":207,"_208":1792,"_210":1793},{"_186":1788,"_24":207,"_208":1789,"_210":1790},{"_186":1785,"_24":207,"_208":1786,"_210":1787},{"_186":1782,"_24":207,"_208":1783,"_210":1784},{"_186":1779,"_24":207,"_208":1780,"_210":1781},"c24ce7265e1d",[],".","c66f9ef7a776",[1770],"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",[1772],"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",[1803,1804,1805,1806,1807],[1797,1798],{"_186":1801,"_24":1079,"_1335":1802},{"_186":1799,"_24":1079,"_1335":1800},"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/",{"_186":1820,"_24":207,"_208":1821,"_210":1822},{"_186":1817,"_24":207,"_208":1818,"_210":1819},{"_186":1814,"_24":207,"_208":1815,"_210":1816},{"_186":1811,"_24":207,"_208":1812,"_210":1813},{"_186":1808,"_24":207,"_208":1809,"_210":1810},"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",[1799],"75%","47cf0e8a331c",[]," to ","d1aebf2378c0",[1801],"30%","adc4d52ad4cd",[],"Major engineering orgs claim that as much as ","f09e9612ef15",[1826],[],{"_186":1827,"_24":207,"_208":1828,"_210":1829},"85de00f41251",[],"Why Code-as-Text Doesn’t Work for Genuine Understanding",[1836,1837,1838,1839,1840,1841,1842,1843,1844,1845,1846,1847,1848,1849,1850,1851,1852,1853,1854,1855,1856,1857,1858,1859,1860,1861,1862,1863,1864,1865,1866,1867,1868,1869,1870,1871,1872,1873,1874,1875,1876,1877,1878,1879,1880,1881,1882,1883,1884,1885,1886,1887,1888,1889,1890,1891,1892,1893,1894,1895,1896,1897,1898,1899,1900,1901,1902,1903,1904,1905,1906,1907,1908,1909,1910,1911,1912,1913,1914,1915,1916,1917,1918,1919,1920,1921,1922,1923],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",{"_186":3043,"_24":188,"_189":3044,"_191":3045,"_193":422},{"_186":3037,"_24":188,"_189":3038,"_191":3039,"_193":221},{"_186":3019,"_24":188,"_189":3020,"_191":3021,"_193":221},{"_186":3012,"_24":188,"_189":3013,"_191":3014,"_193":221},{"_186":3005,"_24":188,"_189":3006,"_1940":1941,"_1942":1943,"_191":3007,"_193":221},{"_186":2998,"_24":188,"_189":2999,"_1940":1941,"_1942":1943,"_191":3000,"_193":221},{"_186":2991,"_24":188,"_189":2992,"_1940":1941,"_1942":1943,"_191":2993,"_193":221},{"_186":2976,"_24":188,"_189":2977,"_191":2978,"_193":221},{"_186":2969,"_24":188,"_189":2970,"_191":2971,"_193":422},{"_186":2962,"_24":188,"_189":2963,"_191":2964,"_193":221},{"_186":2955,"_24":188,"_189":2956,"_191":2957,"_193":221},{"_186":2948,"_24":188,"_189":2949,"_191":2950,"_193":221},{"_186":2941,"_24":188,"_189":2942,"_191":2943,"_193":194},{"_186":2935,"_24":188,"_189":2936,"_191":2937,"_193":194},{"_186":2928,"_24":188,"_189":2929,"_191":2930,"_193":194},{"_186":2921,"_24":188,"_189":2922,"_191":2923,"_193":194},{"_186":2914,"_24":188,"_189":2915,"_191":2916,"_193":194},{"_186":2907,"_24":188,"_189":2908,"_191":2909,"_193":194},{"_186":2900,"_24":188,"_189":2901,"_191":2902,"_193":194},{"_186":2894,"_24":188,"_189":2895,"_191":2896,"_193":194},{"_186":2887,"_24":188,"_189":2888,"_191":2889,"_193":194},{"_186":2880,"_24":188,"_189":2881,"_191":2882,"_193":194},{"_186":2873,"_24":188,"_189":2874,"_191":2875,"_193":194},{"_186":2866,"_24":188,"_189":2867,"_191":2868,"_193":194},{"_186":2859,"_24":188,"_189":2860,"_191":2861,"_193":194},{"_186":2852,"_24":188,"_189":2853,"_191":2854,"_193":194},{"_186":2845,"_24":188,"_189":2846,"_191":2847,"_193":194},{"_186":2838,"_24":188,"_189":2839,"_191":2840,"_193":221},{"_186":2831,"_24":188,"_189":2832,"_191":2833,"_193":221},{"_186":2814,"_24":188,"_189":2815,"_191":2816,"_193":221},{"_186":2811,"_24":1091,"_74":2812},{"_186":2793,"_24":188,"_189":2794,"_191":2795,"_193":221},{"_186":2786,"_24":188,"_189":2787,"_191":2788,"_193":221},{"_186":2779,"_24":188,"_189":2780,"_191":2781,"_193":422},{"_186":2772,"_24":188,"_189":2773,"_191":2774,"_193":221},{"_186":2750,"_24":188,"_189":2751,"_191":2752,"_193":221},{"_186":2720,"_24":188,"_189":2721,"_191":2722,"_193":221},{"_186":2661,"_24":188,"_189":2662,"_191":2663,"_193":221},{"_186":2643,"_24":188,"_189":2644,"_191":2645,"_193":221},{"_186":2636,"_24":188,"_189":2637,"_191":2638,"_193":422},{"_186":2629,"_24":188,"_189":2630,"_191":2631,"_193":221},{"_186":2580,"_24":188,"_189":2581,"_191":2582,"_193":221},{"_186":2573,"_24":188,"_189":2574,"_191":2575,"_193":221},{"_186":2546,"_24":188,"_189":2547,"_191":2548,"_193":221},{"_186":2524,"_24":188,"_189":2525,"_191":2526,"_193":221},{"_186":2497,"_24":188,"_189":2498,"_191":2499,"_193":221},{"_186":2472,"_24":188,"_189":2473,"_191":2474,"_193":221},{"_186":2465,"_24":188,"_189":2466,"_191":2467,"_193":422},{"_186":2458,"_24":188,"_189":2459,"_191":2460,"_193":221},{"_186":2447,"_24":188,"_189":2448,"_191":2449,"_193":221},{"_186":2415,"_24":188,"_189":2416,"_191":2417,"_193":221},{"_186":2408,"_24":188,"_189":2409,"_191":2410,"_193":221},{"_186":2401,"_24":188,"_189":2402,"_191":2403,"_193":194},{"_186":2394,"_24":188,"_189":2395,"_191":2396,"_193":194},{"_186":2387,"_24":188,"_189":2388,"_191":2389,"_193":194},{"_186":2380,"_24":188,"_189":2381,"_191":2382,"_193":194},{"_186":2373,"_24":188,"_189":2374,"_191":2375,"_193":221},{"_186":2362,"_24":188,"_189":2363,"_191":2364,"_193":221},{"_186":2351,"_24":188,"_189":2352,"_191":2353,"_193":221},{"_186":2306,"_24":188,"_189":2307,"_191":2308,"_193":221},{"_186":2299,"_24":188,"_189":2300,"_191":2301,"_193":422},{"_186":2292,"_24":188,"_189":2293,"_191":2294,"_193":221},{"_186":2285,"_24":188,"_189":2286,"_191":2287,"_193":2012},{"_186":2264,"_24":188,"_189":2265,"_1940":1941,"_1942":1943,"_191":2266,"_193":221},{"_186":2254,"_24":188,"_189":2255,"_1940":1941,"_1942":1943,"_191":2256,"_193":221},{"_186":2232,"_24":188,"_189":2233,"_1940":1941,"_1942":1943,"_191":2234,"_193":221},{"_186":2225,"_24":188,"_189":2226,"_191":2227,"_193":221},{"_186":2218,"_24":188,"_189":2219,"_191":2220,"_193":2012},{"_186":2197,"_24":188,"_189":2198,"_1940":1941,"_1942":1943,"_191":2199,"_193":221},{"_186":2187,"_24":188,"_189":2188,"_1940":1941,"_1942":1943,"_191":2189,"_193":221},{"_186":2177,"_24":188,"_189":2178,"_1940":1941,"_1942":1943,"_191":2179,"_193":221},{"_186":2159,"_24":188,"_189":2160,"_191":2161,"_193":221},{"_186":2152,"_24":188,"_189":2153,"_191":2154,"_193":2012},{"_186":2132,"_24":188,"_189":2133,"_1940":1941,"_1942":1943,"_191":2134,"_193":221},{"_186":2111,"_24":188,"_189":2112,"_1940":1941,"_1942":1943,"_191":2113,"_193":221},{"_186":2101,"_24":188,"_189":2102,"_1940":1941,"_1942":1943,"_191":2103,"_193":221},{"_186":2094,"_24":188,"_189":2095,"_191":2096,"_193":221},{"_186":2087,"_24":188,"_189":2088,"_191":2089,"_193":2012},{"_186":2066,"_24":188,"_189":2067,"_1940":1941,"_1942":1943,"_191":2068,"_193":221},{"_186":2045,"_24":188,"_189":2046,"_1940":1941,"_1942":1943,"_191":2047,"_193":221},{"_186":2024,"_24":188,"_189":2025,"_1940":1941,"_1942":1943,"_191":2026,"_193":221},{"_186":2017,"_24":188,"_189":2018,"_191":2019,"_193":221},{"_186":2009,"_24":188,"_189":2010,"_191":2011,"_193":2012},{"_186":1987,"_24":188,"_189":1988,"_1940":1941,"_1942":1943,"_191":1989,"_193":221},{"_186":1965,"_24":188,"_189":1966,"_1940":1941,"_1942":1943,"_191":1967,"_193":221},{"_186":1938,"_24":188,"_189":1939,"_1940":1941,"_1942":1943,"_191":1944,"_193":221},{"_186":1931,"_24":188,"_189":1932,"_191":1933,"_193":221},{"_186":1924,"_24":188,"_189":1925,"_191":1926,"_193":221},"b1115aef2063",[1927],[],{"_186":1928,"_24":207,"_208":1929,"_210":1930},"cc39ba41ca8f",[],"\n","d5296f448a9f",[1934],[],{"_186":1935,"_24":207,"_208":1936,"_210":1937},"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",[1948,1949,1950,1951],"level",1,"listItem","bullet",[1945],{"_186":1946,"_24":1079,"_1335":1947},"b7f12682eee8","https://docs.mubit.ai/concepts/memory-model",{"_186":1961,"_24":207,"_208":1962,"_210":1963},{"_186":1958,"_24":207,"_208":1959,"_210":1960},{"_186":1955,"_24":207,"_208":1956,"_210":1957},{"_186":1952,"_24":207,"_208":1953,"_210":1954},"cf1e8502645e",[]," includes facts, traces, lessons, rules, workflows, and exact archives, with a separate, mutable working state (rather than specifying a particular database backend).","a124afaedcbc",[1946],"memory model","c220cbdd3ded",[]," Mubit’s documented ","a1974bc579f7",[1964],"Store:","strong","1ed494cdb84c",[1971,1972,1973,1974],[1968],{"_186":1969,"_24":1079,"_1335":1970},"03f6430bd2bb","https://docs.mubit.ai/concepts/retrieval-semantics",{"_186":1984,"_24":207,"_208":1985,"_210":1986},{"_186":1981,"_24":207,"_208":1982,"_210":1983},{"_186":1978,"_24":207,"_208":1979,"_210":1980},{"_186":1975,"_24":207,"_208":1976,"_210":1977},"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",[1969],"recall queries","4e86e92fe9b1",[]," The platform uses ","4390006d4807",[1964],"Read:","7f8a916d5d27",[1993,1994,1995,1996],[1990],{"_186":1991,"_24":1079,"_1335":1992},"1db8c126f415","https://docs.mubit.ai/concepts/learning-loop",{"_186":2006,"_24":207,"_208":2007,"_210":2008},{"_186":2003,"_24":207,"_208":2004,"_210":2005},{"_186":2000,"_24":207,"_208":2001,"_210":2002},{"_186":1997,"_24":207,"_208":1998,"_210":1999},"ab2045bc4caf",[]," that attaches outcomes to memory identifiers and uses reflection to derive conditional lessons for validation and promotion.","60a47075eb73",[1991],"learning loop","36a26a2ad674",[]," Mubit uses asynchronous ingestion to classify typed entries. The platform has a ","4a28c2e300ff",[1964],"Write:","23586738a3bd",[2013],[],"h3",{"_186":2014,"_24":207,"_208":2015,"_210":2016},"55a06b57abcf",[],"Mubit","66730d013fe4",[2020],[],{"_186":2021,"_24":207,"_208":2022,"_210":2023},"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",[2030,2031,2032,2033],[2027],{"_186":2028,"_24":1079,"_1335":2029},"0b959dcc7d95","https://docs.cognee.ai/core-concepts/architecture",{"_186":2043,"_24":207,"_208":2044,"_210":1963},{"_186":2040,"_24":207,"_208":2041,"_210":2042},{"_186":2037,"_24":207,"_208":2038,"_210":2039},{"_186":2034,"_24":207,"_208":2035,"_210":2036},"98a98865c12c",[]," combines relational records (for source metadata and provenance), vectors (for semantic search), and a graph (for entities and relationships).","8006ac452a9b",[2028],"architecture","b1af73cfdbec",[]," cognee’s ","3a786e68b48e",[1964],"0c8ba15851e3",[2051,2052,2053,2054],[2048],{"_186":2049,"_24":1079,"_1335":2050},"0d0d58abefcd","https://docs.cognee.ai/core-concepts/main-operations/recall",{"_186":2064,"_24":207,"_208":2065,"_210":1986},{"_186":2061,"_24":207,"_208":2062,"_210":2063},{"_186":2058,"_24":207,"_208":2059,"_210":2060},{"_186":2055,"_24":207,"_208":2056,"_210":2057},"e398196f7eb2",[]," supports different modes according to the evidence/answer required.","a7f71645af37",[2049],"recall interface","b28cafe0f3f1",[]," cognee’s retrieval uses vectors and graph structure. The ","f045c19eb548",[1964],"dcfbde86d09f",[2072,2073,2074,2075],[2069],{"_186":2070,"_24":1079,"_1335":2071},"4171b8028864","https://docs.cognee.ai/core-concepts/main-operations/improve",{"_186":2085,"_24":207,"_208":2086,"_210":2008},{"_186":2082,"_24":207,"_208":2083,"_210":2084},{"_186":2079,"_24":207,"_208":2080,"_210":2081},{"_186":2076,"_24":207,"_208":2077,"_210":2078},"5331b3438bf8",[]," that incorporate useful session information into the permanent graph.","dab6e9a6f8ee",[2070],"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",[1964],"844cbdc9c89b",[2090],[],{"_186":2091,"_24":207,"_208":2092,"_210":2093},"004318b7238a",[],"cognee","a9dca42167a7",[2097],[],{"_186":2098,"_24":207,"_208":2099,"_210":2100},"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",[2104,2105],[],{"_186":2109,"_24":207,"_208":2110,"_210":1963},{"_186":2106,"_24":207,"_208":2107,"_210":2108},"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",[1964],"3417dd3e4105",[2117,2118,2119,2120],[2114],{"_186":2115,"_24":1079,"_1335":2116},"bb5a4597ff40","https://help.getzep.com/graphiti/working-with-data/searching",{"_186":2130,"_24":207,"_208":2131,"_210":1986},{"_186":2127,"_24":207,"_208":2128,"_210":2129},{"_186":2124,"_24":207,"_208":2125,"_210":2126},{"_186":2121,"_24":207,"_208":2122,"_210":2123},"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",[2115],"hybrid search","28d2a41d4e5d",[]," The platform uses a ","788181f131ab",[1964],"200e943d74d2",[2138,2139,2140,2141],[2135],{"_186":2136,"_24":1079,"_1335":2137},"44a39f9fb87c","https://github.com/getzep/graphiti",{"_186":2150,"_24":207,"_208":2151,"_210":2008},{"_186":2148,"_24":207,"_208":2149,"_210":214},{"_186":2145,"_24":207,"_208":2146,"_210":2147},{"_186":2142,"_24":207,"_208":2143,"_210":2144},"bd37a5647c0f",[]," extracts entities and relationships from incoming episodes. The platform can preserve evidence history but allows new evidence to invalidate previous relationships.","646ac358d278",[2136],"Graphiti","2bf68a99ad63",[],"23ab9ad7f341",[1964],"ec1a6ae1873c",[2155],[],{"_186":2156,"_24":207,"_208":2157,"_210":2158},"4d803c10905d",[],"Graphiti and Zep","27fdcaf80cf2",[2165,2166,2167],[2162],{"_186":2163,"_24":1079,"_1335":2164},"39e749b78ea5","https://github.com/mem0ai/mem0",{"_186":2174,"_24":207,"_208":2175,"_210":2176},{"_186":2171,"_24":207,"_208":2172,"_210":2173},{"_186":2168,"_24":207,"_208":2169,"_210":2170},"de1e3d5a2c9a",[]," that differs from the managed architecture described above.)","e5f17f165be9",[2163],"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",[2180,2181],[],{"_186":2185,"_24":207,"_208":2186,"_210":1963},{"_186":2182,"_24":207,"_208":2183,"_210":2184},"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",[1964],"39c6e15e2909",[2190,2191],[],{"_186":2195,"_24":207,"_208":2196,"_210":1986},{"_186":2192,"_24":207,"_208":2193,"_210":2194},"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",[1964],"6625977f2d5a",[2203,2204,2205,2206],[2200],{"_186":2201,"_24":1079,"_1335":2202},"d6a326566adf","https://docs.mem0.ai/core-concepts/memory-evaluation",{"_186":2216,"_24":207,"_208":2217,"_210":2008},{"_186":2213,"_24":207,"_208":2214,"_210":2215},{"_186":2210,"_24":207,"_208":2211,"_210":2212},{"_186":2207,"_24":207,"_208":2208,"_210":2209},"6d3effc41325",[]," processes conversations asynchronously, extracts ADD-only facts, deduplicates them, creates embeddings, and adds entity links and temporal metadata.","60b128b85711",[2201],"extraction pipeline","724c2c891981",[]," Mem0’s ","5a484c730a59",[1964],"813631713196",[2221],[],{"_186":2222,"_24":207,"_208":2223,"_210":2224},"d9f878ff27a0",[],"Mem0's Managed Platform","031f13e60a30",[2228],[],{"_186":2229,"_24":207,"_208":2230,"_210":2231},"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",[2238,2239,2240,2241],[2235],{"_186":2236,"_24":1079,"_1335":2237},"bab345521546","https://www.letta.com/blog/memory-blocks/",{"_186":2251,"_24":207,"_208":2252,"_210":2253},{"_186":2248,"_24":207,"_208":2249,"_210":2250},{"_186":2245,"_24":207,"_208":2246,"_210":2247},{"_186":2242,"_24":207,"_208":2243,"_210":2244},"28d8492c3fe9",[],", which persist labeled, size-limited text. Context repositories represent memory as versioned files, with designated files loaded into the prompt.","2e258a11f020",[2236,1289],"memory blocks","bd33ca7d5b2b",[],"Letta abstracts context window storage as ","3a82d6616947",[1964],"Store: ","f63a7efb1736",[2257,2258],[],{"_186":2262,"_24":207,"_208":2263,"_210":1986},{"_186":2259,"_24":207,"_208":2260,"_210":2261},"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",[1964],"d1ca287a4e65",[2270,2271,2272,2273],[2267],{"_186":2268,"_24":1079,"_1335":2269},"d611832c5bfa","https://www.letta.com/blog/context-repositories/",{"_186":2283,"_24":207,"_208":2284,"_210":2008},{"_186":2280,"_24":207,"_208":2281,"_210":2282},{"_186":2277,"_24":207,"_208":2278,"_210":2279},{"_186":2274,"_24":207,"_208":2275,"_210":2276},"b30686ec4f7d",[]," let memory agents work in isolated Git worktrees and merge their changes.","2162f1f7e868",[2268],"context repositories","7f1857d4cccd",[]," Agents edit persistent memory using tools, with background agents handling consolidation. Letta Code's ","e107ebffb775",[1964],"0893565179df",[2288],[],{"_186":2289,"_24":207,"_208":2290,"_210":2291},"606ad4cb319c",[],"Letta and Letta Code","4f649d90264a",[2295],[],{"_186":2296,"_24":207,"_208":2297,"_210":2298},"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",[2302],[],{"_186":2303,"_24":207,"_208":2304,"_210":2305},"a0bf90cc4808",[],"How Existing Systems Combine These Choices","26e03d89cef4",[2315,2316,2317,2318,2319,2320,2321,2322,2323],[2309,2310],{"_186":2313,"_24":1079,"_1335":2314},{"_186":2311,"_24":1079,"_1335":2312},"c09991de458c","https://arxiv.org/abs/2512.13564","580336df6328","https://www.ibm.com/docs/en/db2/12.1.x?topic=indexes-vector",{"_186":2348,"_24":207,"_208":2349,"_210":2350},{"_186":2345,"_24":207,"_208":2346,"_210":2347},{"_186":2342,"_24":207,"_208":2343,"_210":2344},{"_186":2339,"_24":207,"_208":2340,"_210":2341},{"_186":2336,"_24":207,"_208":2337,"_210":2338},{"_186":2333,"_24":207,"_208":2334,"_210":2335},{"_186":2330,"_24":207,"_208":2331,"_210":2332},{"_186":2327,"_24":207,"_208":2328,"_210":2329},{"_186":2324,"_24":207,"_208":2325,"_210":2326},"2fbdca59fcea",[],".)","c88f2f3993b8",[2311],"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",[1289],"not","f69b2c969e9a",[],"different representations to find similar content. However, vector indexes do ","51dc1c62d004",[1289],"alongside ","4f49328c1072",[]," (specialized data structures that organize embeddings to speed up searches) can sit ","bfff0f664909",[2313,1289],"Vector indexes","65b134718ba7",[1289],"Note: ","459ca3e10e7d",[2354,2355],[],{"_186":2359,"_24":207,"_208":2360,"_210":2361},{"_186":2356,"_24":207,"_208":2357,"_210":2358},"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",[1964],"Entity-relationship graphs.","91041a2f7d82",[2365,2366],[],{"_186":2370,"_24":207,"_208":2371,"_210":2372},{"_186":2367,"_24":207,"_208":2368,"_210":2369},"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",[1964],"Structured records.","f32fca39f7ed",[2376],[],{"_186":2377,"_24":207,"_208":2378,"_210":2379},"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",[2383],[],{"_186":2384,"_24":207,"_208":2385,"_210":2386},"aa317516e315",[],"└── reference_customer_rollout_tracker.md","e6d70c556e93",[2390],[],{"_186":2391,"_24":207,"_208":2392,"_210":2393},"1b55f6b25b56",[],"├── feedback_migration_compatibility.md","9d066e09a773",[2397],[],{"_186":2398,"_24":207,"_208":2399,"_210":2400},"baf15381d1a5",[],"├── MEMORY.md","92b2b15e6ef5",[2404],[],{"_186":2405,"_24":207,"_208":2406,"_210":2407},"593cd17fb5cc",[],"memory/","989c53d303db",[2411],[],{"_186":2412,"_24":207,"_208":2413,"_210":2414},"9091690fe872",[],"Here is an illustrative layout for our example:","586a0cf2fe77",[2424,2425,2426,2427,2428,2429],[2418,2419],{"_186":2422,"_24":1079,"_1335":2423},{"_186":2420,"_24":1079,"_1335":2421},"bf2bc5254bf0","https://code.claude.com/docs/en/memory#storage-location","961db163e8ee","https://code.claude.com/docs/en/skills",{"_186":2444,"_24":207,"_208":2445,"_210":2446},{"_186":2441,"_24":207,"_208":2442,"_210":2443},{"_186":2438,"_24":207,"_208":2439,"_210":2440},{"_186":2435,"_24":207,"_208":2436,"_210":2437},{"_186":2432,"_24":207,"_208":2433,"_210":2434},{"_186":2430,"_24":207,"_208":2431,"_210":1781},"84bc558b7fb4",[],"9b5977a7be50",[2420],"default memory is local to the machine","8411ca9c5db5",[],", but the ","3dbe26891ce6",[2422,1289],"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",[1964],"Text files and summaries.","221d0943bbbe",[2450,2451],[],{"_186":2455,"_24":207,"_208":2456,"_210":2457},{"_186":2452,"_24":207,"_208":2453,"_210":2454},"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",[1964],"Raw logs and artifacts.","4ce4eb451d82",[2461],[],{"_186":2462,"_24":207,"_208":2463,"_210":2464},"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",[2468],[],{"_186":2469,"_24":207,"_208":2470,"_210":2471},"788c87082178",[],"Storage: Representation, Backend, and Indexes","7d7e734abb9d",[2481,2482,2483,2484],[2475,2476],{"_186":2479,"_24":1079,"_1335":2480},{"_186":2477,"_24":1079,"_1335":2478},"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",{"_186":2494,"_24":207,"_208":2495,"_210":2496},{"_186":2491,"_24":207,"_208":2492,"_210":2493},{"_186":2488,"_24":207,"_208":2489,"_210":2490},{"_186":2485,"_24":207,"_208":2486,"_210":2487},"856e85a5c3db",[],", supporting continuity within a session while treating the selection of durable lessons for future sessions as a separate decision.","66110f9314ae",[2477],"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",[2479,1289],"Compaction","1d284f0bdd3e",[2500,2501,2502,2503,2504,2505],[],{"_186":2521,"_24":207,"_208":2522,"_210":2523},{"_186":2518,"_24":207,"_208":2519,"_210":2520},{"_186":2515,"_24":207,"_208":2516,"_210":2517},{"_186":2512,"_24":207,"_208":2513,"_210":2514},{"_186":2509,"_24":207,"_208":2510,"_210":2511},{"_186":2506,"_24":207,"_208":2507,"_210":2508},"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",[1289],"how retrieval works","17b0e32a878e",[]," initiates retrieval is separate from ","9c22117de48b",[1289],"Who","d277a5a24136",[],": ","8d32937852d7",[1289],"About retrieval and memory overhead","baec8a07be2b",[2530,2531,2532,2533],[2527],{"_186":2528,"_24":1079,"_1335":2529},"0f84a55baee9","https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents",{"_186":2543,"_24":207,"_208":2544,"_210":2545},{"_186":2540,"_24":207,"_208":2541,"_210":2542},{"_186":2537,"_24":207,"_208":2538,"_210":2539},{"_186":2534,"_24":207,"_208":2535,"_210":2536},"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",[2528],"Claude Code combines upfront context with on-demand retrieval","de9d369d54d7",[]," Another approach is to let agents use tools to investigate. For example, ","5db73706b4b4",[1964],"Let the model retrieve memory as needed.","3a95128b53ea",[2549,2550,2551,2552,2553,2554],[],{"_186":2570,"_24":207,"_208":2571,"_210":2572},{"_186":2567,"_24":207,"_208":2568,"_210":2569},{"_186":2564,"_24":207,"_208":2565,"_210":2566},{"_186":2561,"_24":207,"_208":2562,"_210":2563},{"_186":2558,"_24":207,"_208":2559,"_210":2560},{"_186":2555,"_24":207,"_208":2556,"_210":2557},"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",[1289],"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",[1289],"before","8406b79505b0",[]," Another approach runs retrieval ","7061a6e00d28",[1964],"Have the harness retrieve and assemble context.","da3b59765f18",[2576],[],{"_186":2577,"_24":207,"_208":2578,"_210":2579},"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",[2589,2590,2591,2592,2593,2594,2595,2596,2597,2598],[2583,2584],{"_186":2587,"_24":1079,"_1335":2588},{"_186":2585,"_24":1079,"_1335":2586},"2fe7d1e0fb7c","https://code.claude.com/docs/en/memory#claudemd-vs-auto-memory","6b5816824a71","https://code.claude.com/docs/en/memory#how-it-works",{"_186":2626,"_24":207,"_208":2627,"_210":2628},{"_186":2623,"_24":207,"_208":2624,"_210":2625},{"_186":2620,"_24":207,"_208":2621,"_210":2622},{"_186":2617,"_24":207,"_208":2618,"_210":2619},{"_186":2614,"_24":207,"_208":2615,"_210":2616},{"_186":2611,"_24":207,"_208":2612,"_210":2613},{"_186":2608,"_24":207,"_208":2609,"_210":2610},{"_186":2605,"_24":207,"_208":2606,"_210":2607},{"_186":2602,"_24":207,"_208":2603,"_210":2604},{"_186":2599,"_24":207,"_208":2600,"_210":2601},"f04028db5006",[]," between persistent instructions vs. auto memory.)","72cea0fbced3",[2585],"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",[1289],"write ","a7474d19202d",[]," with ","d9ba8f226466",[1289],"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",[2587],"loads MEMORY.md at conversation start","351d6bf3467b",[]," One approach is to keep a small set of instructions or pointers within context (Claude Code ","81afda08f68e",[1964],"Include designated records automatically.","0d2f5b0ff899",[2632],[],{"_186":2633,"_24":207,"_208":2634,"_210":2635},"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",[2639],[],{"_186":2640,"_24":207,"_208":2641,"_210":2642},"978e8914fa68",[],"Read Policy: What Reaches the Model?","28f6819109be",[2646,2647,2648,2649],[],{"_186":2658,"_24":207,"_208":2659,"_210":2660},{"_186":2655,"_24":207,"_208":2656,"_210":2657},{"_186":2653,"_24":207,"_208":2654,"_210":2335},{"_186":2650,"_24":207,"_208":2651,"_210":2652},"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",[1289],"2635523039da",[]," These write policies are ","fcd21325e323",[1289],"Note:","5861fa445fa5",[2672,2673,2674,2675,2676,2677,2678,2679,2680,2681,2682,2683],[2664,2665,2666],{"_186":2670,"_24":1079,"_1335":2671},{"_186":2668,"_24":1079,"_1335":2669},{"_186":2667,"_24":1079,"_1335":2269},"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",{"_186":2717,"_24":207,"_208":2718,"_210":2719},{"_186":2714,"_24":207,"_208":2715,"_210":2716},{"_186":2711,"_24":207,"_208":2712,"_210":2713},{"_186":2708,"_24":207,"_208":2709,"_210":2710},{"_186":2705,"_24":207,"_208":2706,"_210":2707},{"_186":2702,"_24":207,"_208":2703,"_210":2704},{"_186":2699,"_24":207,"_208":2700,"_210":2701},{"_186":2696,"_24":207,"_208":2697,"_210":2698},{"_186":2693,"_24":207,"_208":2694,"_210":2695},{"_186":2690,"_24":207,"_208":2691,"_210":2692},{"_186":2687,"_24":207,"_208":2688,"_210":2689},{"_186":2684,"_24":207,"_208":2685,"_210":2686},"91234870a812",[]," for documented examples of using background consolidation to update agent memory.","938fcb5e36a8",[2667],"Letta Code's context repositories","b37a5929e4ab",[]," (deciding whether new facts should update an existing memory). See ","5e3b514ea686",[1289],"consolidation","cdd413341cbd",[],") from ","f07d4da484cd",[2668],"candidate facts","00c26c30913f",[]," (used to identify ","ce486f61cec6",[1289],"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",[2670,1289],"extractor","c729269fd4b9",[]," Another approach is processing recorded sessions independently of the agent with an ","097ba3fb5001",[1964],"Use a separate extraction pipeline.","b0388f909a33",[2726,2727,2728,2729,2730,2731],[2723],{"_186":2724,"_24":1079,"_1335":2725},"86369451d645","https://code.claude.com/docs/en/memory#auto-memory",{"_186":2747,"_24":207,"_208":2748,"_210":2749},{"_186":2744,"_24":207,"_208":2745,"_210":2746},{"_186":2741,"_24":207,"_208":2742,"_210":2743},{"_186":2738,"_24":207,"_208":2739,"_210":2740},{"_186":2735,"_24":207,"_208":2736,"_210":2737},{"_186":2732,"_24":207,"_208":2733,"_210":2734},"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",[1289],"worth ","a377237b09e8",[]," selectively saves user info, feedback, context, and references, skipping facts recoverable from code or Git history. The difference is what is ","ea35d09d64e9",[2724],"auto memory","d0750afe804b",[]," Claude Code's ","a3d9b929d5ec",[1964],"Let the active agent write a memory.","d809f9f408ed",[2756,2757,2758,2759],[2753],{"_186":2754,"_24":1079,"_1335":2755},"f5a4020c4fb9","https://code.claude.com/docs/en/how-claude-code-works#work-with-sessions",{"_186":2769,"_24":207,"_208":2770,"_210":2771},{"_186":2766,"_24":207,"_208":2767,"_210":2768},{"_186":2763,"_24":207,"_208":2764,"_210":2765},{"_186":2760,"_24":207,"_208":2761,"_210":2762},"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",[2754],"records messages, tool calls, and results","3c08d4da140a",[]," Claude Code ","d53459c4b154",[1964],"Automatically capture execution.","ea5d8e984d9d",[2775],[],{"_186":2776,"_24":207,"_208":2777,"_210":2778},"6cedcb2c785b",[],"A write policy determines what gets retained, who interprets it, and when the result becomes available. These three are most common:","ccf4da4770a3",[2782],[],{"_186":2783,"_24":207,"_208":2784,"_210":2785},"424c91921778",[],"Write Policy: What Becomes Memory?","d676ce5bde85",[2789],[],{"_186":2790,"_24":207,"_208":2791,"_210":2792},"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",[2799,2800,2801],[2796],{"_186":2797,"_24":1079,"_1335":2798},"e04bb9c15fd2","https://x.com/himanshustwts/status/2038924027411222533",{"_186":2808,"_24":207,"_208":2809,"_210":2810},{"_186":2805,"_24":207,"_208":2806,"_210":2807},{"_186":2802,"_24":207,"_208":2803,"_210":2804},"921f830b5e17",[1289],". 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",[2797,1289],"himanshu on X, March 31, 2026","083a41cf889e",[1289],"Source: ","ad95b655bc4d",{"_1062":2813,"_24":1064},"image-f8350e46e885d0b2fa3ac0a247b0dbef59722d0b-1200x720-jpg","6196452f03db",[2819,2820,2821],[2817],{"_186":2818,"_24":1079,"_1335":2798},"b6eb1cd451c2",{"_186":2828,"_24":207,"_208":2829,"_210":2830},{"_186":2825,"_24":207,"_208":2826,"_210":2827},{"_186":2822,"_24":207,"_208":2823,"_210":2824},"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",[2818],"himanshu's walkthrough of Claude Code's memory architecture","3a123eb6cc1c",[],"I came across ","7abfa2abb045",[2834],[],{"_186":2835,"_24":207,"_208":2836,"_210":2837},"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",[2841],[],{"_186":2842,"_24":207,"_208":2843,"_210":2844},"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",[2848],[],{"_186":2849,"_24":207,"_208":2850,"_210":2851},"904ce784cc6b",[],")","33e2a8d52b3b",[2855],[],{"_186":2856,"_24":207,"_208":2857,"_210":2858},"7680b5ad5684",[],"outcome_t","f4d41d0ffe96",[2862],[],{"_186":2863,"_24":207,"_208":2864,"_210":2865},"80d216e10ccb",[],"action_t,","7f8cc217016d",[2869],[],{"_186":2870,"_24":207,"_208":2871,"_210":2872},"9d5241992604",[],"observation_t,","79b91869e2e3",[2876],[],{"_186":2877,"_24":207,"_208":2878,"_210":2879},"76a995443776",[],"M_t,","f082f460dd34",[2883],[],{"_186":2884,"_24":207,"_208":2885,"_210":2886},"66125ac80fbe",[],"M_t+1 = F(","d3f764a17d9c",[2890],[],{"_186":2891,"_24":207,"_208":2892,"_210":2893},"72a3b7698b71",[],"","cbc2ed7ad74a",[2897],[],{"_186":2898,"_24":207,"_208":2899,"_210":2851},"2816fd2f0683",[],"e6879dfd579c",[2903],[],{"_186":2904,"_24":207,"_208":2905,"_210":2906},"2611c7cb1e9f",[],"observation_t","f22ddf871918",[2910],[],{"_186":2911,"_24":207,"_208":2912,"_210":2913},"affdd49b21ac",[],"context_t,","0b328b527a81",[2917],[],{"_186":2918,"_24":207,"_208":2919,"_210":2920},"2a217f0d1f95",[],"working_state_t,","fbf8e2258ec2",[2924],[],{"_186":2925,"_24":207,"_208":2926,"_210":2927},"ac90fae6dc2e",[],"instructions,","d90d632afed0",[2931],[],{"_186":2932,"_24":207,"_208":2933,"_210":2934},"f9eaeeeb18fd",[],"action_t = model(","bccb7605f45d",[2938],[],{"_186":2939,"_24":207,"_208":2940,"_210":2893},"4b7a59ffa024",[],"62cc0bf02726",[2944],[],{"_186":2945,"_24":207,"_208":2946,"_210":2947},"c20486f8ce97",[],"context_t = R(query_t, M_t, B)","8a7574d4e676",[2951],[],{"_186":2952,"_24":207,"_208":2953,"_210":2954},"6a62e755d8ee",[],"A memory-augmented agent adds persistent state M_t, a write policy F, a retrieval policy R, and a context budget B:","73cf6663ed99",[2958],[],{"_186":2959,"_24":207,"_208":2960,"_210":2961},"28413b799a65",[],"action_t = model(instructions, recent_messages, observation_t)","02f0e6bf9487",[2965],[],{"_186":2966,"_24":207,"_208":2967,"_210":2968},"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",[2972],[],{"_186":2973,"_24":207,"_208":2974,"_210":2975},"52f17d0d3b1a",[],"Memory as a Stateful Control Loop","cb4932e369b2",[2979,2980,2981],[],{"_186":2988,"_24":207,"_208":2989,"_210":2990},{"_186":2985,"_24":207,"_208":2986,"_210":2987},{"_186":2982,"_24":207,"_208":2983,"_210":2984},"e314c861bd9f",[]," the model, in the harness layer that controls the agent loop.","d84a78243562",[1289],"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",[2994],[],{"_186":2995,"_24":207,"_208":2996,"_210":2997},"18e76a78164b",[],"It cannot coordinate with other agents if each action begins with an isolated view of the organization","72565cf8678f",[3001],[],{"_186":3002,"_24":207,"_208":3003,"_210":3004},"744b85306b2c",[],"Learning can't happen without memory, and an agent cannot learn if the outcome of one attempt disappears before the next","a0eeaffd296d",[3008],[],{"_186":3009,"_24":207,"_208":3010,"_210":3011},"0f74cd6c9885",[],"An agent cannot pursue a goal over months if it forgets why the goal/sub-goal matters","7139f92c87c1",[3015],[],{"_186":3016,"_24":207,"_208":3017,"_210":3018},"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",[3025,3026,3027],[3022],{"_186":3023,"_24":1079,"_1335":3024},"f72dcfadb07b","https://www.heavybit.com/library/article/long-horizon-agents-from-order-takers-to-outcome-owners",{"_186":3034,"_24":207,"_208":3035,"_210":3036},{"_186":3031,"_24":207,"_208":3032,"_210":3033},{"_186":3028,"_24":207,"_208":3029,"_210":3030},"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",[3023],"Part 1","dade4a80a130",[],"In ","829b03b465e4",[3040],[],{"_186":3041,"_24":207,"_208":3042,"_210":2893},"224565565806",[],"16e0361b4ce2",[3046],[],{"_186":3047,"_24":207,"_208":3048,"_210":3049},"991570e6c69f",[],"Why Memory Is A Key Piece of the Long-Horizon Puzzle",[3056,3057,3058,3059,3060,3061,3062,3063,3064,3065,3066,3067,3068,3069,3070,3071,3072,3073,3074,3075,3076,3077,3078,3079,3080,3081,3082,3083,3084,3085,3086,3087,3088,3089,3090,3091,3092,3093,3094,3095,3096,3097,3098,3099,3100,3101,3102,3103,3104,3105,3106],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",{"_186":3661,"_24":188,"_189":3662,"_191":3663,"_193":422},{"_186":3643,"_24":188,"_189":3644,"_191":3645,"_193":221},{"_186":3625,"_24":188,"_189":3626,"_191":3627,"_193":221},{"_186":3622,"_24":1091,"_74":3623},{"_186":3608,"_24":188,"_189":3609,"_191":3610,"_193":221},{"_186":3601,"_24":188,"_189":3602,"_191":3603,"_193":422},{"_186":3583,"_24":188,"_189":3584,"_191":3585,"_193":221},{"_186":3560,"_24":188,"_189":3561,"_191":3562,"_193":221},{"_186":3545,"_24":188,"_189":3546,"_191":3547,"_193":221},{"_186":3538,"_24":188,"_189":3539,"_191":3540,"_193":221},{"_186":3531,"_24":188,"_189":3532,"_191":3533,"_193":422},{"_186":3524,"_24":188,"_189":3525,"_191":3526,"_193":221},{"_186":3509,"_24":188,"_189":3510,"_191":3511,"_193":221},{"_186":3502,"_24":188,"_189":3503,"_191":3504,"_193":221},{"_186":3495,"_24":188,"_189":3496,"_191":3497,"_193":221},{"_186":3488,"_24":188,"_189":3489,"_191":3490,"_193":221},{"_186":3481,"_24":188,"_189":3482,"_191":3483,"_193":221},{"_186":3458,"_24":188,"_189":3459,"_191":3460,"_193":221},{"_186":3451,"_24":188,"_189":3452,"_191":3453,"_193":422},{"_186":3444,"_24":188,"_189":3445,"_191":3446,"_193":221},{"_186":3437,"_24":188,"_189":3438,"_191":3439,"_193":221},{"_186":3419,"_24":188,"_189":3420,"_191":3421,"_193":221},{"_186":3412,"_24":188,"_189":3413,"_191":3414,"_193":194},{"_186":3405,"_24":188,"_189":3406,"_191":3407,"_193":422},{"_186":3379,"_24":188,"_189":3380,"_191":3381,"_193":221},{"_186":3364,"_24":188,"_189":3365,"_191":3366,"_193":221},{"_186":3350,"_24":188,"_189":3351,"_191":3352,"_193":221},{"_186":3335,"_24":188,"_189":3336,"_191":3337,"_193":221},{"_186":3328,"_24":188,"_189":3329,"_191":3330,"_193":221},{"_186":3321,"_24":188,"_189":3322,"_191":3323,"_193":221},{"_186":3314,"_24":188,"_189":3315,"_191":3316,"_193":422},{"_186":3299,"_24":188,"_189":3300,"_191":3301,"_193":221},{"_186":3292,"_24":188,"_189":3293,"_191":3294,"_193":221},{"_186":3261,"_24":188,"_189":3262,"_191":3263,"_193":221},{"_186":3254,"_24":188,"_189":3255,"_191":3256,"_193":221},{"_186":3247,"_24":188,"_189":3248,"_191":3249,"_193":422},{"_186":3240,"_24":188,"_189":3241,"_191":3242,"_193":221},{"_186":3233,"_24":188,"_189":3234,"_191":3235,"_193":221},{"_186":3226,"_24":188,"_189":3227,"_191":3228,"_193":221},{"_186":3219,"_24":188,"_189":3220,"_191":3221,"_193":221},{"_186":3212,"_24":188,"_189":3213,"_191":3214,"_193":422},{"_186":3194,"_24":188,"_189":3195,"_191":3196,"_193":221},{"_186":3163,"_24":188,"_189":3164,"_191":3165,"_193":221},{"_186":3156,"_24":188,"_189":3157,"_191":3158,"_193":221},{"_186":3149,"_24":188,"_189":3150,"_191":3151,"_193":221},{"_186":3142,"_24":188,"_189":3143,"_191":3144,"_193":221},{"_186":3135,"_24":188,"_189":3136,"_191":3137,"_193":194},{"_186":3128,"_24":188,"_189":3129,"_191":3130,"_193":422},{"_186":3121,"_24":188,"_189":3122,"_191":3123,"_193":221},{"_186":3114,"_24":188,"_189":3115,"_191":3116,"_193":221},{"_186":3107,"_24":188,"_189":3108,"_191":3109,"_193":221},"3e791b59e928",[3110],[],{"_186":3111,"_24":207,"_208":3112,"_210":3113},"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",[3117],[],{"_186":3118,"_24":207,"_208":3119,"_210":3120},"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",[3124],[],{"_186":3125,"_24":207,"_208":3126,"_210":3127},"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",[3131],[],{"_186":3132,"_24":207,"_208":3133,"_210":3134},"b2d7e327e929",[],"The Future of Orchestration (From the Terminal or Otherwise)","526f25bd891d",[3138],[],{"_186":3139,"_24":207,"_208":3140,"_210":3141},"e081e6730529",[1289],"“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",[3145],[],{"_186":3146,"_24":207,"_208":3147,"_210":3148},"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",[3152],[],{"_186":3153,"_24":207,"_208":3154,"_210":3155},"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",[3159],[],{"_186":3160,"_24":207,"_208":3161,"_210":3162},"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",[3166,3167,3168,3169,3170,3171,3172],[],{"_186":3191,"_24":207,"_208":3192,"_210":3193},{"_186":3188,"_24":207,"_208":3189,"_210":3190},{"_186":3185,"_24":207,"_208":3186,"_210":3187},{"_186":3182,"_24":207,"_208":3183,"_210":3184},{"_186":3179,"_24":207,"_208":3180,"_210":3181},{"_186":3176,"_24":207,"_208":3177,"_210":3178},{"_186":3173,"_24":207,"_208":3174,"_210":3175},"03ff68661418",[]," with the fix removed.”","71d9a0844a69",[1289],"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",[1289],"worse","89f7daac29b3",[]," is ","c68efdd5d660",[1289],"a test that can't fail","e7a2992b03e7",[],"“The other half of that lesson is that ","2c567aa855f5",[3200,3201,3202],[3197],{"_186":3198,"_24":1079,"_1335":3199},"f924a1482c30","https://github.com/Gaurav-Gosain/tuitest",{"_186":3209,"_24":207,"_208":3210,"_210":3211},{"_186":3206,"_24":207,"_208":3207,"_210":3208},{"_186":3203,"_24":207,"_208":3204,"_210":3205},"807d067e16ed",[],".”","e6b851d76bb4",[3198],"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",[3215],[],{"_186":3216,"_24":207,"_208":3217,"_210":3218},"b09cc780d360",[],"Lessons Learned: Testing, Ownership, Docs, Community","28c61fd9c130",[3222],[],{"_186":3223,"_24":207,"_208":3224,"_210":3225},"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",[3229],[],{"_186":3230,"_24":207,"_208":3231,"_210":3232},"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",[3236],[],{"_186":3237,"_24":207,"_208":3238,"_210":3239},"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",[3243],[],{"_186":3244,"_24":207,"_208":3245,"_210":3246},"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",[3250],[],{"_186":3251,"_24":207,"_208":3252,"_210":3253},"3a4e64e47a2d",[],"The Future of the Terminal: Long Horizon and Org-Wide Projects","3cc74079e800",[3257],[],{"_186":3258,"_24":207,"_208":3259,"_210":3260},"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",[3264,3265,3266,3267,3268,3269,3270],[],{"_186":3289,"_24":207,"_208":3290,"_210":3291},{"_186":3286,"_24":207,"_208":3287,"_210":3288},{"_186":3283,"_24":207,"_208":3284,"_210":3285},{"_186":3280,"_24":207,"_208":3281,"_210":3282},{"_186":3277,"_24":207,"_208":3278,"_210":3279},{"_186":3274,"_24":207,"_208":3275,"_210":3276},{"_186":3271,"_24":207,"_208":3272,"_210":3273},"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",[1289],"more","47dfd88b7b4f",[]," than it has but never ","1a9fe7c73af4",[1289],"less","860c3089b9db",[],", so an agent can give a helper ","837e89a39a8a",[1289],"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",[3295],[],{"_186":3296,"_24":207,"_208":3297,"_210":3298},"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",[3302,3303,3304],[],{"_186":3311,"_24":207,"_208":3312,"_210":3313},{"_186":3308,"_24":207,"_208":3309,"_210":3310},{"_186":3305,"_24":207,"_208":3306,"_210":3307},"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",[1289],"specifically run through the project itself","c439e261144e",[],"Gosain points out that the project has built-in identity management for jobs ","bd5192f93adf",[3317],[],{"_186":3318,"_24":207,"_208":3319,"_210":3320},"bf4caaeb1172",[],"Managing Agents Via Terminal: Identity and Loopholes","df24faceca60",[3324],[],{"_186":3325,"_24":207,"_208":3326,"_210":3327},"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",[3331],[],{"_186":3332,"_24":207,"_208":3333,"_210":3334},"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",[3338,3339,3340],[],{"_186":3347,"_24":207,"_208":3348,"_210":3349},{"_186":3344,"_24":207,"_208":3345,"_210":3346},{"_186":3341,"_24":207,"_208":3342,"_210":3343},"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",[1289],"deny","9c3818dff01a",[],"“In practice, that means only you can answer an agent's permission prompt. Another agent can ","d4a501c3e7f1",[3353,3354,3355],[],{"_186":3361,"_24":207,"_208":3362,"_210":3363},{"_186":3358,"_24":207,"_208":3359,"_210":3360},{"_186":3356,"_24":207,"_208":3357,"_210":3205},"79b7b12169fb",[],"fb8f26544685",[1289],"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",[3367,3368,3369],[],{"_186":3376,"_24":207,"_208":3377,"_210":3378},{"_186":3373,"_24":207,"_208":3374,"_210":3375},{"_186":3370,"_24":207,"_208":3371,"_210":3372},"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",[1289],"comparing","ba5da87edf7b",[],"“The part I care about most is ","f65d4a7bf402",[3385,3386,3387,3388,3389],[3382],{"_186":3383,"_24":1079,"_1335":3384},"3ed3816cd07a","https://github.blog/ai-and-ml/github-copilot/what-are-git-worktrees-and-why-should-i-use-them/",{"_186":3402,"_24":207,"_208":3403,"_210":3404},{"_186":3399,"_24":207,"_208":3400,"_210":3401},{"_186":3396,"_24":207,"_208":3397,"_210":3398},{"_186":3393,"_24":207,"_208":3394,"_210":3395},{"_186":3390,"_24":207,"_208":3391,"_210":3392},"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",[3383],"worktree","d13353fe843a",[]," functionality to parallelize task execution across multiple agents, each in its own ","46ba47a14fde",[1289],"fan","ca141d2fe81c",[],"The creator notes that he specifically built a ","8fc105268f08",[3408],[],{"_186":3409,"_24":207,"_208":3410,"_210":3411},"291bd897e642",[],"Failstates, Agent Reliability, and Learning Job-to-Job","80403ff8cd0a",[3415],[],{"_186":3416,"_24":207,"_208":3417,"_210":3418},"8d2aebf3369f",[1289],"“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",[3425,3426,3427],[3422],{"_186":3423,"_24":1079,"_1335":3424},"eb4ac0b4ccd2","https://www.cnbc.com/2026/06/16/anthropics-fable-shutdown-is-a-big-moment-for-open-source-ai.html",{"_186":3434,"_24":207,"_208":3435,"_210":3436},{"_186":3431,"_24":207,"_208":3432,"_210":3433},{"_186":3428,"_24":207,"_208":3429,"_210":3430},"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",[3423],"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",[3440],[],{"_186":3441,"_24":207,"_208":3442,"_210":3443},"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",[3447],[],{"_186":3448,"_24":207,"_208":3449,"_210":3450},"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",[3454],[],{"_186":3455,"_24":207,"_208":3456,"_210":3457},"a7aeb5be88b4",[],"Where Terminal-Level Orchestration Fits in the Big Picture","fa331092dd99",[3461,3462,3463,3464,3465],[],{"_186":3478,"_24":207,"_208":3479,"_210":3480},{"_186":3475,"_24":207,"_208":3476,"_210":3477},{"_186":3472,"_24":207,"_208":3473,"_210":3474},{"_186":3469,"_24":207,"_208":3470,"_210":3471},{"_186":3466,"_24":207,"_208":3467,"_210":3468},"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",[1289],"‘Is this actually the right change?’","14d1aa383e6c",[]," will mostly get automated away. ","459f8e217e2c",[1289],"‘Can this agent run this command?’","f60396b582a3",[],"“I also think the bottleneck moves over time. ","cedcc0d17912",[3484],[],{"_186":3485,"_24":207,"_208":3486,"_210":3487},"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",[3491],[],{"_186":3492,"_24":207,"_208":3493,"_210":3494},"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",[3498],[],{"_186":3499,"_24":207,"_208":3500,"_210":3501},"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",[3505],[],{"_186":3506,"_24":207,"_208":3507,"_210":3508},"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",[3512,3513,3514],[],{"_186":3521,"_24":207,"_208":3522,"_210":3523},{"_186":3518,"_24":207,"_208":3519,"_210":3520},{"_186":3515,"_24":207,"_208":3516,"_210":3517},"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",[1289],"attention","e98c0abcd22c",[],"“What breaks first, at least for me, is ","7b661806cc77",[3527],[],{"_186":3528,"_24":207,"_208":3529,"_210":3530},"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",[3534],[],{"_186":3535,"_24":207,"_208":3536,"_210":3537},"467d878226a1",[],"Where Do Agent-Plus-Human-in-the-Loop Systems Break?","724b27bc87e5",[3541],[],{"_186":3542,"_24":207,"_208":3543,"_210":3544},"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",[3548,3549,3550],[],{"_186":3557,"_24":207,"_208":3558,"_210":3559},{"_186":3554,"_24":207,"_208":3555,"_210":3556},{"_186":3551,"_24":207,"_208":3552,"_210":3553},"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",[1289],"which agent needed him","01c30f4c4a0b",[],"As Gosain started running several agents in parallel, arranging panes stopped being the problem. Knowing ","c8f8b537e2f0",[3563,3564,3565,3566,3567],[],{"_186":3580,"_24":207,"_208":3581,"_210":3582},{"_186":3577,"_24":207,"_208":3578,"_210":3579},{"_186":3574,"_24":207,"_208":3575,"_210":3576},{"_186":3571,"_24":207,"_208":3572,"_210":3573},{"_186":3568,"_24":207,"_208":3569,"_210":3570},"bcfd2c3fa106",[],"”","b4a8ddc05174",[1289],"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",[1289],"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",[3589,3590,3591],[3586],{"_186":3587,"_24":1079,"_1335":3588},"66e4fafd0d2e","https://github.com/tmux/tmux/wiki",{"_186":3598,"_24":207,"_208":3599,"_210":3600},{"_186":3595,"_24":207,"_208":3596,"_210":3597},{"_186":3592,"_24":207,"_208":3593,"_210":3594},"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",[3587],"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",[3604],[],{"_186":3605,"_24":207,"_208":3606,"_210":3607},"49637b6e2cd9",[],"From Multiple Terminals to Multi-Agent","98f9ba5b1089",[3614,3615],[3611],{"_186":3612,"_24":1079,"_1335":3613},"f04212d3515e","https://www.tuios.dev",{"_186":3619,"_24":207,"_208":3620,"_210":3621},{"_186":3616,"_24":207,"_208":3617,"_210":3618},"9b006e84ff8b",[3612,1289],"tuios.dev","51417f32ca16",[1289],"TUIOS helps users orchestrate agents across multiple terminals. Image courtesy ","d4790a7a95ba",{"_1062":3624,"_24":1064},"image-31b56f7b3c41c15c73d7b410e290c2f05c7c5ede-1600x478-jpg","fa72a247de6c",[3631,3632,3633],[3628],{"_186":3629,"_24":1079,"_1335":3630},"3e20d7e69a4e","https://github.com/Gaurav-Gosain/tuios",{"_186":3640,"_24":207,"_208":3641,"_210":3642},{"_186":3637,"_24":207,"_208":3638,"_210":3639},{"_186":3634,"_24":207,"_208":3635,"_210":3636},"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",[3629],"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",[3649,3650,3651],[3646],{"_186":3647,"_24":1079,"_1335":3648},"52e334b19f82","https://en.wikipedia.org/wiki/Terminal_multiplexer",{"_186":3658,"_24":207,"_208":3659,"_210":3660},{"_186":3655,"_24":207,"_208":3656,"_210":3657},{"_186":3652,"_24":207,"_208":3653,"_210":3654},"932af58ea732",[]," to give each agent their own separate pane that segregates and preserves the work of each across sessions.","ae81e2f7bd1b",[3647],"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",[3664],[],{"_186":3665,"_24":207,"_208":3666,"_210":3667},"24ae57392a2a",[],"Do Windows Matter Less Than Limited Human Attention?",{"_1212":3699,"_96":3700,"_1091":3673,"_1216":1204,"_51":3701,"_14":3702,"_63":3703,"_1223":3677},{"_1212":3686,"_96":3687,"_1091":3673,"_1216":1202,"_51":3688,"_14":3689,"_63":3690,"_1223":3677},{"_1212":3671,"_96":3672,"_1091":3673,"_1216":1200,"_51":3674,"_14":3675,"_63":3676,"_1223":3677},[3678],"53:20","https://cdn.sanity.io/images/50q6fr1p/production/3df09eb0c5e853d8b193072234d9485ab211374e-3000x3000.jpg","2026-08-20T07:10:00.000Z","lab-notes/ep-4-the-new-big-data-of-inference-with-junchen-jiang","Lab Notes Ep. #4, The New Big Data of Inference with Junchen Jiang","podcastEpisode",{"_186":3679,"_24":188,"_189":3680,"_191":3681,"_193":221},"045e7251e9db",[3682],[],{"_186":3683,"_24":207,"_208":3684,"_210":3685},"47a21c16c04d",[],"On episode 4 of Lab Notes, Amir Zohrenejad speaks with Junchen Jiang about why KVCache may be better understood as reusable, AI-native data rather than a temporary inference optimization. They explore how LMCache and CacheBlend can reduce redundant computation, move context across distributed inference systems, and help support increasingly complex AI agents. The conversation also covers multimodal workloads, open-source infrastructure, and the future of AI systems research.",[3691],"41:13","2026-09-02T07:10:00.000Z","lab-notes/ep-5-the-voice-ai-turing-test-with-aoden-teo","Lab Notes Ep. #5, The Voice AI Turing Test with Aoden Teo",{"_186":3692,"_24":188,"_189":3693,"_191":3694,"_193":221},"d8d3c0bd9e91",[3695],[],{"_186":3696,"_24":207,"_208":3697,"_210":3698},"43855ba2a734",[],"In episode 5 of Lab Notes, Amir Zohrenejad speaks with Aoden Teo, co-founder and CEO of Miso Labs, about what it will take for voice AI to become truly humanlike. They explore the limitations of today's speech-to-text and text-to-speech pipelines, the promise of full-duplex speech models, and the technical challenges around audio tokenization, turn-taking, emotion, reasoning, and real-time inference.",[3704],"01:02:12","2026-10-22T07:15:00.000Z","lab-notes/ep-6-why-ai-needs-to-argue-with-itself-with-james-evans","Lab Notes Ep. #6, Why AI Needs to Argue with Itself with James Evans",{"_186":3705,"_24":188,"_189":3706,"_191":3707,"_193":221},"155560d05cc5",[3708],[],{"_186":3709,"_24":207,"_208":3710,"_210":3711},"e08a50142832",[],"On episode 6 of Lab Notes, Amir Zohrenejad speaks with James Evans about the surprising finding that reasoning models simulate a society rather than a single brain. They dig into sycophancy and why users trust AI less but perform better when it has a point of view, the Computational Life experiments that evolved self-replicating programs from a primordial soup of code, and James’s argument that diverse novel ideas are about to become the scarcest resource on the planet.",{"_186":3713,"_24":188,"_189":3714,"_191":3715,"_193":221},"exlhs0",[3716],[],{"_186":3717,"_24":207,"_208":3718,"_210":71},"exlhs0s",[],"ua","Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; ClaudeBot/1.0; +claudebot@anthropic.com)","hideBanner",{"_3741":3742,"_3743":3744,"_3745":3746,"_3747":-5},"ENV",{"_3725":3726,"_3727":3728,"_3729":3730,"_3731":3732,"_3733":3734,"_3735":3736,"_3737":3738,"_3739":3726,"_3740":3726},"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",{"_3859":3860,"_3861":3862,"_1078":3863},"globalConfig",{"_72":3856},"header",{"_3748":3749,"_3750":3751},"skills","announcementBar",{"_3848":3849,"_3850":-5,"_3851":31,"_3852":3853,"_210":3854,"_90":3855},"headerNavLinks",[3752,3753,3754,3755,3756],{"_186":3829,"_3758":3830,"_3759":3831,"_1079":3832,"_3762":12,"_63":3833},{"_186":3814,"_3758":3815,"_3759":3816,"_1079":3817,"_3762":12,"_63":3818},{"_186":3796,"_3758":3797,"_3759":3798,"_1079":3799,"_3762":12,"_63":3800},{"_186":3764,"_3758":3765,"_3759":3766,"_1079":3767,"_3762":12,"_63":3768},{"_186":3757,"_3758":-5,"_3759":3760,"_1079":3761,"_3762":31,"_63":3763},"3702498a502c","dropdown","label","Contact","/press","mobileOnly","Navigate to contact page","4bec0dd665d4",[3769,3770,3771,3772,3773,3774],"Library","/library","Navigate to library page",{"_186":3792,"_3759":3793,"_1079":3794,"_63":3795},{"_186":3788,"_3759":3789,"_1079":3790,"_63":3791},{"_186":3784,"_3759":3785,"_1079":3786,"_63":3787},{"_186":3781,"_3759":3782,"_1079":3783,"_63":3782},{"_186":3778,"_3759":3779,"_1079":3780,"_63":3779},{"_186":3775,"_3759":3776,"_1079":3777,"_63":3776},"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",[3801,3802,3803],"Community","/community","Navigate to community page",{"_186":3811,"_3759":3812,"_1079":3813,"_63":3812},{"_186":3808,"_3759":3809,"_1079":3810,"_63":3809},{"_186":3804,"_3759":3805,"_1079":3806,"_63":3807},"23077fca0e2f","DevGuild","/devguild","Navigate to devguild page","1fb7c9c465d0","Advisors","/community/advisors","3a177efc4649","Events","/events","90b68b861c06",[3819,3820],"Portfolio","/portfolio","Navigate to portfolio page",{"_186":3825,"_3759":3826,"_1079":3827,"_63":3828},{"_186":3821,"_3759":3822,"_1079":3823,"_63":3824},"e50fa43ad2e7","Jobs","/jobs","Portfolio Company Jobs","d481fcbbd3af","Spotlights","/portfolio/spotlights","Navigate to spotlights page","a8d641babed2",[3834,3835,3836],"About","/about","Navigate to about page",{"_186":3844,"_3759":3845,"_1079":3846,"_63":3847},{"_186":3840,"_3759":3841,"_1079":3842,"_63":3843},{"_186":3837,"_3759":3838,"_1079":3761,"_63":3839},"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":1091,"_74":3857},{"_1062":3858,"_24":1064},"image-32fdfae4fd13b8f9a3c5bf6b13a635068cb0ecbb-1332x529-png","emailSubscription",{"_979":3929,"_3930":3931},"footerColumns",[3895,3896,3897],[3864,3865,3866,3867,3868],{"_186":3890,"_3870":3891,"_90":3892},{"_186":3885,"_3870":3886,"_90":3887},{"_186":3880,"_3870":3881,"_90":3882},{"_186":3875,"_3870":3876,"_90":3877},{"_186":3869,"_3870":3871,"_90":3872},"639729864c2d","icon",{"_24":1094,"_74":3873},"https://www.tiktok.com/@heavybitindustries",{"_1062":3874,"_24":1064},"image-48ac9473e30341c4bee9f6634b77b6f8f1dd6b2f-849x980-webp","a100f6448036",{"_24":1094,"_74":3878},"https://www.instagram.com/heavybitindustries/",{"_1062":3879,"_24":1064},"image-b38b7ee9ccd05f61ca43d5b29fe7058f822f87eb-512x512-webp","fae1cc0df5c6",{"_24":1094,"_74":3883},"https://www.youtube.com/c/heavybit",{"_1062":3884,"_24":1064},"image-6c548d2e815d1fb3b3077c9b8be1ec251e1bc418-512x512-png","5348b028f0d41c3dac62b3dd2b5531ee",{"_24":1094,"_74":3888},"https://www.linkedin.com/company/heavybit",{"_1062":3889,"_24":1064},"image-01bd92eb06873bdc30b1cb5f9c78b1ef5a9559d4-512x512-png","ee025c96397c",{"_24":1094,"_74":3893},"https://twitter.com/heavybit",{"_1062":3894,"_24":1064},"image-9d618da0948299428b29e5d499af3fa5369a24d2-20x20-svg",{"_186":3921,"_3899":3831,"_1205":3922},{"_186":3909,"_3899":3798,"_1205":3910},{"_186":3898,"_3899":3766,"_1205":3900},"ea618aa34610","heading",[3901,3902,3903],{"_186":3906,"_3759":3907,"_1079":3767,"_63":3908},{"_186":3905,"_3759":3785,"_1079":3786,"_63":3787},{"_186":3904,"_3759":3776,"_1079":3777,"_63":3776},"96a97ceba3a2","e1673433290e","c82a64edc4dd","Blog","Navigate to browse on library page","a584a6acf07b",[3911,3912,3913,3914],{"_186":3920,"_3759":3812,"_1079":3813,"_63":3812},{"_186":3919,"_3759":3805,"_1079":3806,"_63":3807},{"_186":3918,"_3759":3822,"_1079":3823,"_63":3824},{"_186":3915,"_3759":3916,"_1079":3917,"_63":3916},"e22bfa6b2330","Inclusion","/community/diversity-and-inclusion","feb038ce46a1","e2d5641ef7de","4358699f4ca4","30a87ef6e7c1",[3923,3924,3925],{"_186":3928,"_3759":3816,"_1079":3817,"_63":3818},{"_186":3927,"_3759":3845,"_1079":3846,"_63":3847},{"_186":3926,"_3759":3838,"_1079":3761,"_63":3839},"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"]

The Long-Horizon Stack

Speakers