Ep. #8, One Human Plus Agents with Scott Breitenother

Dori Wilson: Hi, I'm Dori Wilson. Welcome to another episode of Data Renegades.

CL Kao: Hi, I'm CL, CEO and founder of Recce and your host of Data Renegades. Today our guest is Scott Breitenother. Scott is the co-founder of Kilo Code and fast growing open source AI coding agent.

But before Kilo, he built Brooklyn Data Company from his apartment into a 100-person consultancy before successful exit and founded LocallyOptimistic, one of the go to communities for data ML leaders.

Hello Scott, welcome to the podcast.

Scott Breitenother: Hello. That was a great introduction. I feel like you're a good confidence boost. I should start every day with a CL introduction. I think I should record that.

CL: You'll definitely get a recording. Haha!

Scott: Perfect. Amazing.

Dori: Yeah. Everyone should start with a hype man. CL is awesome in that way.

Scott: He's very good at that. Yeah, I totally agree.

Dori: Yeah.

CL: Yeah. I remember I first met Scott in one of the DBT conferences, like maybe 2023, before I moved to SF and then he was like, where's the wi-fi password? And we kept running into each other at conferences.

Scott: I'm glad you remember that. I do remember randomly meeting you. I didn't know it was for an embarrassing where's the wi-fi password question. I think your answer was probably--

CL: "It's right on the badge."

Scott: Right on the badge, wasn't it? Thank you. Right where it always is.

CL: All Right, cool. So Scott, can you like take us back to the beginning? I know you're not doing data nowadays but you started your career doing a lot of data stuff and then even built this really big consultancy. So what problem first pulled you into the data space?

Scott: Oh man. I mean I've found I've always been just analytical. But I think the big transition I had was I started my career as a management consultant. So just like crushing Excel and PowerPoint. I don't like to brag about many things, but at that point I was spectacular at Excel and PowerPoint. Now I'm just like, at best, okay.

But what I loved was the transition from that analytical semi-technical background I had in the management consulting world to jumping into my first data job at Casper, the direct to consumer mattress company.

So I had a first assignment, it was to like understand the customer base at Casper. And it was just like using that analytical foundation that I had honed in the mean streets of Excel and then bringing that to the modern data stack, SQL. I mean that was just such an amazing inflection point and just got me hooked.

CL: All right, so from consultancy and then managing consultancy and then working through Casper and then like putting that skill into a more engineered practice.

Scott: Yeah, I mean, a scalable technology. I mean, again, I think Excel is awesome and it's like the first programming language.

Dori: It will never die.

Scott: It'll never die. Y ou might not be writing formulas directly anymore, you might be using AI to do 90% of it. Like that's another thing. But like, I mean it's a very logical kind of semi-structured programming language and I think it's a really great on ramp.

Some people stop at Excel and that's totally fine. But I think it's a really great on ramp to then get to like the more scalable engineering esque, you know, world where you're using SQL and Python and cloud data warehouses and things like that. It's a bit of a gateway drug.

Dori: Yeah. And like queries in Excel and Google sheets, once you actually get into SQL, your life is so much easier, I have found. But I was never an Excel fan girl. It was always, I have to use this. I know some basic lookups and that was it.

Scott: I mean maybe we should do a part two where Scott just shows Dori the wonderful sides of Excel. I just think you just haven't, you haven't seen its finer sides. I think you'd fall in love. It's like the Paris of Microsoft Office. It's just you can't not fall in love.

Dori: Yeah, but is there a lot competing with it in Microsoft Office?

Scott: You know what, I'm not going to put a horse in this race. I'm not making any enemies in the Microsoft Office team.

CL: But you know, prior to Excel, it was really like the spreadsheet concept from like Lotus and then also going back to, I think something associated with like HyperCard. So that was really, like what you say, a gateway drug, as in a way for people to represent logic and transformation. And basically if you decompose that it is still a DAG. Right? And then if you get like a wrench, it is like applying DAG to a bunch of data.

Scott: I don't know if you're just saying that to humor me, but I'll agree 100%. I totally agree. It's the original DAG.

CL: Yeah. I had a chance to work on an open source project called EtherCalc, which is basically like a very small online collaborative Excel thing. And it was a super fun thing to work on.

So maybe talk a little bit about the data stack for a bit. I know you're not in data anymore, we will get into that. But from your perspective, what's actually broken in the modern data stack?

Scott: Yeah, I mean, I wouldn't say broken, but I think there needs to be more change.

I feel like the modern data stack has kind of stagnated in the last few years.

I think you've seen more features in all the products, but I don't know. And maybe this is just the nostalgia of the early years when stuff was just really broken and so there was no way to go but up. But the products have just all gotten quite complex and more feature rich without necessarily feeling like--

I haven't had that magical moment in data in a while. And I think now that I'm more on the AI side, I feel like every day is magic. And I think, yes, AI is genuinely magic.

And seeing the speed at the kind of engineers at Kilo move using AI is phenomenal. But I do actually think it isn't just that AI is generally magic, I think it's just the early days of modern data stack felt magical. It was that unlock, that first time you used dbt, the first time you move from a Read Replica on Postgres to Redshift. And the first time you moved from Redshift and not having to worry about like fixed core sizes and going to Snowflake.

I just haven't had that magical moment and I haven't seen that magical moment in a few years.

Dori: Do you think it's-- because they have more features and it's really cool. Like one of my favorite tools that's popped up is Hex, which is just basically a souped up Jupyter notebook in the best way. Do you think it's just that we haven't built anything new or do you think a lot of the problems are solved in some of these ways?

Scott: Yeah, I think it might be that, like I could just be the old man. It's like I remember when I used to have to walk uphill in the snow both ways to make a dashboard and now, you know, we've solved 90% of it and it's just maturing and maybe that's it. And I do think like Hex is a phenomenal tool.

I remember actually when Barry was like just ideating on Hex, you know, in the early days, it was so funny how hard it was to share and collaborate on Jupyter notebooks. And yeah, I forgot what was the package that everybody was using for kind of shared open Jupyter notebooks. It just wasn't easy.

Dori: Oh, I was like downloading those PDFs to share to PMs so they could operate GitHub. They weren't going to do that.

Scott: And then Hex comes out and just makes it magical and seamless and it just like Appleifies it. But I also think even Hex. Like, I think Hex has found these really cool moments to weave AI into the experience. But even that, I mean, Hex is in and of itself maturing as a product.

And so I think maybe Dori, you're absolutely right that we're just in a phase where we're not waking up every single day to see some sort of magical innovation in the data space. And maybe again, nostalgic people like me just need to accept that you don't have to walk uphill both ways in the snow anymore.

Dori: Yeah, or you still do. But I think, I don't know, like talking to people and being in this space and being in data infra, it's not so much the infra and the technical stuff, I think, I mean, there's cost savings. Like you said there's nothing groundbreaking necessarily that we've seen yet, you know, knock on wood.

But how do you interpret the data? How do you work with it? How do you story-tell, like, the actual functional use of the data?

Scott: Totally.

Dori: That is still, I think, difficult. And that is where LLMs are exciting, but they hallucinate so much.

Scott: But there hasn't been an aha moment. And I think we're getting closer. And again, I'm putting on my "Scott from Kilo" hat for a moment, like, just even seeing how the LLMs have just gotten better in the last couple months. You know, like, Opus 4.5 felt like a breakthrough on the coding side and even on the engineering side.

It does feel like we're starting to get to some sort of inflection point where the LLMs can take on messier, more complex and more autonomous workloads. And who knows, maybe it is time to see the LLMs take on storytelling. I think you're definitely starting to see LLMs help with visualization, building some sort of semantic layer behind the scenes. I think there's this last mile of storytelling that I think once the LLMs get there, that's going to be a magical moment.

Dori: Yeah. And a scary moment for a lot of us who work in data.

Scott: Yeah, I'd like to think you've got to be an optimist in data. You see a lot of things. I think the data team sees the most amount of information about their organization. If you're not an optimist, you're not going to survive very long.

And so I'd like to think the field of data is a field of optimists. And I myself feel like, in general, an optimist. So I don't think people are going to be put out of a job by AI.

I mean, I think sounds cliche. You might be put out of a job by someone who knows AI because they're more productive and efficient and know how to use the tools. But I think if you know the tools, you're not going anywhere. But the expectation will be instead of telling one story, you're going to be telling 20 stories and just the productivity expectations are going to go up.

Dori: Yep.

CL: Yeah, we had an episode with Ben Stancil and then he's been talking about maybe this whole data thing is a fad. Right? Everything is vibesy, and the storytelling is kind of like trying to tell the story. And then the trend is upish. Right? So he's been putting out thoughts on there, but I guess I'm with you. Like more optimistic and hopefully not just locally, but globally optimistic for where we're heading.

Scott: Yeah.

CL: Okay, so for whatever is happening in the data world, but given your experience working with so many clients, what's a tool or workflow that kind of quietly serves teams, but that's kind of under advocated, or people don't talk about it, but if you adopt this thing, then things are great.

Scott: Yeah. I mean this is maybe non technical, but I think just asking questions and being okay with asking questions and having humility. Like, I think one of the things that I think a lot of people that joined Brooklyn Data early on were potentially surprised by was that they didn't have to like sound smart internally.

Like, it was okay saying I don't know and asking other people a question about something. I think in too many organizations-- And I realize I probably took this question in a very different direction than you intended, but I'm going with it. I'm committing.

Dori: I love it.

Scott:

I think in so many organizations you are kind of forced to act like you know everything, even when you don't. And I don't think we learn, I don't think we help each other, I don't think it serves us well if we always put that face on.

And so listen, you should ask smart questions. I'm a big fan of, in general, always asking multiple choice questions instead of free text questions. Instead of like, what is this? It's like, you know, I did some research, is it this or that?

But I think being honest and open with what you know and don't know I think, you know, that was again, this more on the softer side. But I think that served us really, really well at Brooklyn Data and helped us actually, ironically learn faster. If you're honest about what you do and don't know, it's a lot easier to fill in those gaps.

Dori: Yeah. Get at the "so what?" and what is the real problem there? Yeah, it reminds me of the meme that's like the normal curve where, at the bottom, the beginner is saying "I don't know." And the top is like always having the answer. And on the other side, where it's like the experienced person is like, "I don't know" again, I think that's the curve that I see with a lot of seniority.

CL: The Jedi, I don't know.

Dori: Yeah. Yeah, exactly. Exactly.

Scott: Yeah.

It's the confidence to admit you don't know, which I think is actually a pretty cool moment, when you can get there professionally.

Dori: Yeah. How did you build that culture internally? Because that's a difficult thing to build, I think, especially for data people, where we want to have the answer.

Scott: Yeah, I think it kind of came from necessity, to be honest. Like, I found a lot of cultural elements at Brooklyn Data just came out of just our origin story. You know, we were remote during COVID. Most of us hadn't even met in person before. And so the only way you can really work together is to work really loudly and just be honest and transparent.

Dori: Yeah.

Scott: And I don't know what that specific moment was and maybe it's just kind of my unnatural comfort just saying, I don't know, that maybe rubbed off on everybody. But I think from an early, early on we said, hey, don't ask in DMs, ask in a channel. Get it out there, if there's a question. Odd are if you don't know, someone else also doesn't know. And we will benefit as an organization.

And then I think the ultimate cherry on top is like, okay, cool, let's document that somewhere. I don't think I did this well early on at Brooklyn Data, but we got this later on. And then at Kilo we're doing it all the time is like Slack is not a enduring, long lasting documentation center. We, in fact, at Kilo expire, you know, messages after 90 days. Like at Brooklyn Data we did it at 13 months.

But the idea is you shouldn't be looking up something in Slack. Slack is for real time communication. And then when there's something that is worthy of documentation, you should put it in whatever, the kind of the central repository of knowledge. At Brooklyn Data that was Jira. At Kilo, we have a Docusaurus repository site that we have internal access to.

CL: That reminds me of the GitLab handbook. Right? It's very handbook first and then it's now your co-founder.

Scott: Yeah, it's fun. So now, you know what CL was saying is my co-founder at Kilo, Sid Sijbrandij founded GitLab. And what's kind of very full circle is when Brooklyn Data started, we were remote from day one.

So started in I guess August, September of 2018. And the only vocal remote organizations were GitLab and Buffer. And I think we learned, like we voraciously consumed all of the, the kind of knowledge and kind of documentation that those two organizations put out and so we learned so much about how to operate his organization from GitLab and now getting a chance to partner with Sid every day, it's very cool.

I feel like longtime listener, first time caller, constantly. Sid is just this person I've been reading about and kind of consuming his philosophy on working. And now I get to partner with him and work with him every day.

CL: Wow, that must feel amazing.

Dori: Yeah. How did y'all meet up and start founding?

Scott: Yeah, it's funny, we were introduced by a mutual acquaintance or a mutual friend, Emilie Schario, who everybody knows in the data community. She's kind of a former GitLabber, kind of in the data space and I guess co-moderator admin of LocallyOptimistic with me. And just right place, right time, you know, Sid was looking for a co-founder and I was ripe to start my next thing after my mini retirement after selling Brooklyn data.

Dori: Awesome. You know, between founding, both of you, you know, multi-time founders, what is something that you took from the first time into this time that you think has really paid off for you?

Scott: Oh man, I certainly don't have the stamina I had 10 years ago. So I hope to heck I'm wiser. I'd like to think I am. But one of the things that I've really been trying to be very conscious of is and I'm not consistently good at doing this, I'll be very transparent about it, is at Brooklyn Data I jumped into every Slack conversation.

Like every topic I was like every problem I would jump in and solve. And I think it's because like I knew the most because I was there since day one. I was working the most hours I had, you know, maybe the most skin in the game and that it helped us move fast. Like we'd get to a decision fast, I would jump in, make a decision, move on.

What also happened, however, side effect is the organization did not develop the muscle to make decisions without me and I really noticed that as the organization got longer when I couldn't be there in, in every conversation or maybe I was still trying to or attempting to, that the autonomy, the decision making muscle had not developed in the organization and I had kind of put myself as like this critical link, maybe just let's call a spade a spade, a bottleneck into things.

And so I'm trying to be very conscious to you know, let a Slack conversation go. Like the temptation is always there to jump in. And what I'll actually do is I will often bookmark a conversation. And my playbook now is when I see a thread that normally I would have jumped into, I will right click it and I will select to get notified about replies.

So even though I'm not mentioned, I'll get the replies, and then I will set a reminder to come back in three hours to see if it was replied to. And it actually ends up being, in the short term, more work. But I think in the long term, it's been really cool to see.

It's not like the Brooklyn Data team wasn't capable of doing that, but I never gave them a chance. And so I'm really trying to be thoughtful on just not falling into that same trap and letting the work build that muscle. And it's been really great to see.

CL: Yeah, yeah, I can totally resonate with that. Because, you know, as a founder, you're building three things, right? The product, the business, and the organization. And then organization is kind of your best leverage, like getting people to have that while getting the right people having the right ways of working to build.

Scott: Yeah, you got Dori, right? Your secret weapon.

Dori: Yeah, it's great. I love my ego being fluffed right here. It's fantastic.

Scott: You know, we should all hype each other this whole call. We'll start every Monday with this.

CL: Amazing.

Dori: Yeah, no, the vibes are great. Starting off with optimism, because that is not something I associate typically with data. So I'm loving the vibes from y'all.

Scott: What?

Dori: I think, coming from academia, that's my background in economics, where a lot of times you're throwing cold water on people and be like, actually, no, what you thought is not the answer. Or you're just trying to poke holes in arguments, that's a lot of academics.

Scott: Oh, you must be real fun at a party.

Dori: Haha. Years in the industry have helped.

Scott: Okay, good, good, good. Significantly.

CL: Okay, so I want to segue a little bit, like, in between data and AI. Right? So first, I guess, what are your thoughts about data democratization? Did that actually happen, or are we waiting for AI to make it happen as we talk about storytelling and all that AI can do?

Scott: Yeah, I mean, that's a great question. I mean, I would say that I've never been the biggest fan of everybody in an organization should know SQL. I feel like you put something, you know, in the primary query versus a CTE, you could change the answer pretty dramatically. And, like, you know, SQL itself isn't hard, but some of that kind of nuance of how to structure a SQL query is challenging.

And so I was okay with having a little friction between someone having access to the data is. I felt like pure data democracy was gonna end up in bad outcomes over time. I've maybe learned that like I'd rather have people using data more and I'm more okay with some of the risk of kind of wrong numbers or outcomes.

At least, like the better solution is to maybe build relationships with folks so that they know to reach out to the data team if it's a big decision. But like who am I to stand in the way of someone using data?

But I think we're kind of getting to the point where I may be able to have my cake and eat it too. We can do data democracy and have more guardrails of, say, having AI help versus someone having to write a SQL query from scratch.

Like, are the AI models perfect now? No, but I think if I, as a centralized data team, if we can curate the data model well enough and give enough context to the AI model to help a semi-technical or non-technical user answer their own question, that's phenomenal.

You always have the agency problem if you're giving someone else the ability to do something. There's risk, but I think the risk is minimizing and, optimist here, listen, sometimes you're gonna stub your toe, but I mean like that's what happens when you're running fast. It happens.

Dori: Yeah. Is there a stack you like for helping people right now ask those questions. Is there a setup you have or a workflow?

Scott: You know, I don't think there's a perfect solution. I think you're seeing all the BI tools integrate AI features in, you're seeing even kind of like other players that haven't traditionally been in the data space. Like Dagster has Compass. We haven't found that perfect solution. And the question is, is it going to be a BI specific tool or is it going to be a generalist AI agent or harness that helps you do this?

I don't know what the answer is. We're still in that space, but it's very exciting. I don't think any of the solutions are perfect, but it gives me the kind of confidence that in the next 18 to 24 months we're going to be in a great spot. But also BI tools are sticky and so people are still using Cognos and other things. So like it might take many years for this stuff to trickle into all organizations.

Dori: Yeah.

CL: So Scott, what I'm hearing about what you mentioned about data democracies, maybe it was or the way we frame it was a means to the end. Right? It's like we still want people to be curious and with their agency, like finding the answers, but not necessarily with everyone writing SQL.

But now having those questions surface in an either very well organized data team or agent system that could answer those questions. Right?So is that another form of data democracy or is it completely different?

Scott: I don't know. I mean, who knows? But by the way, I love how you synthesized what I was saying over the course of five minutes into something very smart and succinct in like 15 seconds. What CL said, that's what I should have said.

I think it's just an evolution. Each new technology and the modern data stack, the MPP database, the ability to massively analyze data, it unlocks a cool new superpower. All of these technologies, they're not robots replacing us, they're exoskeletons that make us better, faster, stronger. And so this is just like the latest, greatest exoskeleton that magnifies the innate human ability.

Summarize that, CL. That was good.

CL: I totally feel that, personally, I feel it's almost like 20 years ago. I'm coding every day, playing with all these things and building random stuff. So totally, exoskeleton.

Scott: Mhm.

CL: So how do you see AI really reshaping the role of data engineer or a data team in general?

Scott: Yeah, it's the exoskeleton.

Like all organizations and all knowledge workers, you're going to be expected to do more.

And I don't think it's--

CL: Parkinson's law. Haha.

Scott: Yeah, I don't think it's replacing people's jobs. So like I always like to make metaphors that I have no experience in and don't even know they're right. But I'll make this one. I'd like to think like a stock broker or a stock analyst, like in the 50s was covering like five stocks and then you know, in the 70s it was 20 stocks and then 50 stocks and now it's maybe 200 stocks. Like with AI you have to cover 10,000 stocks.

And I think it's just going to be the exact same thing where you know, any knowledge worker and data is an example of a knowledge worker, you're going to be expected to do a lot more. And so we have a data team at Kilo of one, Pedro, I used to work with him at Brooklyn Data crushes it, one-man army. And he does more than a team of three or four would have a few years ago. And yes Pedro, if you're listening, you're amazing.

But I also think it's because he embraces the technology he embraces using-- He uses Kilo to write all the, you know, Kilo plus the DBT MCP server and he's cruising. I think that's just what the role is going to be.

We have a philosophy within Kilo, it might sound bad, but like we're very anti-collaboration. And it's funny Dori, I saw you at a PostHog mug and I think that's like. I give a lot of credit to the PostHog folks. They had a great blog post about this.

We have one developer, one marketer, one data person, own a feature end to end. They can collaborate with other team members of course if they need it. Like, you know, it's not like you can't talk with anybody. But I think too often humans use collaboration as like a safety blanket.

Like you know, we need to have a check in meeting or you need to look at my code or let me put this, this document on my methodology out there for two weeks and get feedback. At Kilo we just move fast. It's a team member plus a team of agents. Just you know, that's the unit.

And so you know, we have team members like Suresh owns code reviews, develop that end to end from conception to production, the Kilo feature. Pedro owns our data stack end to end. We don't need five Pedros. We have a team of five. It's just Pedro plus four agents. And that's kind of how we operate and we see high performing teams starting to operate.

CL: So what I'm hearing is autonomous, single focus. Because now experienced team members can work with agents and that essentially is a whole team. Right? And then they are kind of collaborating with the agent more than kind of their peers as you set that boundary. But of course you still encourage that if necessary, right?

Scott: Yeah, the team still exists but it's not five humans, it's one human and four agents.

Dori: Yeah, the end to end data person. I've been seeing that a lot like you, you know the threads on Reddit and people talking about LinkedIn job postings. I think you're seeing that a lot more of kind of the full stack that can do everything. The data analyst, the analytics engineer, and the data engineer.

Scott: Yeah.

Dori: And just roll with that all the way through.

Scott: 100%. I mean, that's the full stack data person, I think we've really embraced it and just Pedro joined right before Christmas and so he's what, two months in or something like that, we have a full data stack and he's just ripping through, kind of answering people's questions, but also adding big incremental features to the data set constantly.

And I see this even as a CEO at Kilo, just deciding what we do next, what we build next. The calculus has changed so much because you can do it so quickly. You know, you can go from, you know, we want to build a feature to it being done in a week.

We can go from having like, you know, we want to ingest these three data sources and be reporting on it to it being done in two days. Like, the speed is just kind of pretty wild and that feels magical.

CL: Yeah, Yeah, I can totally relate to that. Kilo is like forked from like Roo Code. Right? I played with the earlier version of that about a year ago and things have changed so much.

Scott: Yeah, we launched over this past summer. So as a company, we started in March and the product launched over the summer. And so we're less than a year old as a company and, well less than a year old as a feature. I mean, like, yeah, we're a superset of Cline and Roo, but we built a lot of really cool features. Check it out.

What would have changed since you first experienced Kilo was, you know, initially was a VS code extension and just a really good VS code extension. Now we've built the end to end agentic engineering platform so you can do VS code, CLI, JetBrains AppBuilder, which is kind of like Lovable, code reviews, Kilo for Slack, parallel agents so you can operate multiple agents at the same time.

It's pretty wild because we have this saying at Kilo, which is called Kilo Speed, which is just when you're just flying. I mean it's like when, when you're just like, I know everything's going right, the luck is on your side, the winds are in the right direction, the technology is working. And I think that's what we aspire to be operating at at all times is Kilo Speed.

All those features that I just described, they've been built in the last few months. I mean in most companies that would be a 6 to 12 to maybe 18 month roadmap.

Dori: Yep.

Scott:

I mean we're ripping through features every day. That's what happens when you get a really great full stack individual and you wrap them in an AI exoskeleton and that's what we're building here at Kilo is that kind of awesome AI exoskeleton.

So I guess meta, the Kilo engineers are using Kilo to build Kilo every day. And it's pretty magical to watch.

Dori: And the best type of dog fooding too.

Scott: Exactly. And a lot of our features came out that way. Like parallel agents just came from-- It's like, oh, shoot, I have to run one agent at a time. Like that's the bottleneck. And so. Okay, cool. Let's build the infrastructure to run multiple agents. When you are your target customer, it's great.

But also we have the kind of the structure where all of our engineers who own the products actually engage directly with the customers. W e have a customer support team, but there's no kind of layer in between. And Suresh who owns code reviews is interfacing directly with all the customers, the users, to get feedback to make code reviews better every single day.

And speaking of magical moments. Like our users are really kind of blown away by when they give it feedback and like by end of the day it's in production, it's like, hey by the way, can you reload the, you know, download the latest version of the extension? The fix is in there. And they're just like mind blown because they've never seen the pace of rapid iteration that we're doing.

CL: Yeah. Wow.I can see this is kind of like the future of how things will work. Have you run into any-- Although you said it's anti-collaboration, but this is still kind of like you're shipping things like different features. And so have you run into new problems like this way that things got conflict, or the feature is incompatible. What are kind of the learnings there?

Scott: So what we'll say is a human reviews every PR before it goes out. I mean it's an AI assisted review, but we are not YOLOing code straight into our code base. And we made it very clear from day one that you are accountable for any code that you commit, regardless of whether you physically typed in the characters yourself or you used AI to generate them you're accountable.

But what we found is really interesting is there's always like this one person that maybe everybody had do the code review. And what we've done is we've used AI to automate, I would say 95% of our code reviews. And so, you know, we're kind of getting to a point where in some ways AI is better.

And I guess what I mean is as code bases get more and more complex, what's more likely to catch a conflict or an error? The person that's been there 10 years and has that hard earned knowledge of the code base and the why this happened, or the AI agent that is parsing every single line of code and kind of interrogating it all from every single angle to see if there are conflicts? I think the human is still going to win every once in a while. Just going to find, ah, this is something the AI hasn't, etc.

But the models are getting so good that it's getting to the point where AI is actually better.

And what you're starting to see is kind of by having a number of checks. So as I'm a developer, I'm checking periodically to do a local code review. We launched local code reviews on Kilo, I think two weeks ago. And then I push it live, push it to Git, make a PR and then there's another automated code review. And then what we're also working on is our security reviewer and kind of more proactive reviews that, that are just kind of continuing to look for issues autonomously.

And so I think having that kind of multiple levels of automated checks, they're exoskeletons again, they're not going to replace a human, but they're kind of adding a lot of value and in some cases surpassing what a human could do on their own and allowing us to move at Kilo speed and to give autonomy because we, we know that AI is going to be looking for broader conflicts in the system.

CL: Wow. Okay, so all this exoskeleton enabling things to move at Kilo speed and then what are some of this learning you got from this that you can give advice to data teams to move at Kilo speed as well?

Scott: Yeah, I think a few learnings were we've gotten to a point where writing code is actually quite, quite fast. And then you look at what are the other bottlenecks in an organization. It is often all the bureaucracy, all the kind of unnecessary decision gates, all the kind of comfort blankets that we put in as a society and as an organization to make people feel comfortable making decision and moving forward.

What we've done is we've tried to ruthlessly hunt those down and get rid of them. And so when you join at Kilo, your first day isn't like let's tour the office and meet people. It's like, you're expected to have code shipped by the end of day, and you're expected to have a fully working MVP feature by the end of the week. And it's expected to be in prod the next week.

There is no warmup period. And you kind of see this journey that an engineer goes like, at first they're like, oh, like this intimidation. And then it's like, um, can I really? Like, it's like, you know, then it's a validation. And then once you're like, yeah, go write some code, then it's just like, you see them fly.

Dori: Yeah.

Scott: Nobody wants to be checking, getting permission all the time. You know, listen, yes, they should know that they can ask a question. Like, you're always allowed to ask a question. But also we've got a out of the box specialist agent called Ask that you can ask right at Kilo and ask the code base. And so you don't even have to ask a human. You don't have to wait for someone to respond. You just go.

And so I love seeing this kind of light that kind of goes in someone's eyes when they realize, like, wow, actually I can do this. And our developers at Kilo aren't working 24 hours a day. They're not working seven days a week. They're just really productive because they're using agents and we're not making them ask for permission and check in every five minutes.

Dori: Love it.

CL: Okay. Should all the data teams do that?

Scott: Yeah, I mean, I think data team members have always been quite autonomous. Like, you know, I think we've just had to do it just by the nature of just having to be scrappy and get things together. But I think yes is the answer that, like, we need to just take that autonomy to the next level and just ship.

The trend that I've seen in my data experience is we've always been just a couple years behind the engineers in adapting some of the Git modern engineering kind of software development life cycle.

I think the data team can't afford to be two years behind. You need to be using AI to generate the code. You need to be using AI to help you do the analysis. You need to get that exoskeleton going now.

And so yeah Cl, to answer your question, it's like yes, we need to be making kind of full stack data folks that can move faster and use their exoskeletons.

CL: Okay, that's great. Okay, so a couple more questions before we go into our lightning rounds. I guess you're on the frontier of all these coding agent things. This is amazing but fast forward one year. What's going to feel like just really laughably outdated about how we do things especially with data or like the coding or the way that we operate today?

Scott: Yeah.

I think you're going to see a big shift to longer and longer autonomous agentic workloads tackling more and more complex tasks. And so maybe right now you're letting an agent run for 5, 10, 15, 20 minutes. You're going to let an agent run for hours.

You know, you're going to let them, you know, give it a mission of doing your entire data stack migration, giving it requirements, heck even having it define its own requirements, its own testing and acceptance criteria and then just running at it. And not just one agent, but multiple agents in parallel.

And I think the kind of sequential or hands on nature that we're currently using to manage agents, I think that's going to feel a bit dated. And you know, this data person, the developer, heck any knowledge worker, their job is going to be kicking off long running autonomous agentic workloads, having those agentic workloads actually flag the human when they have need a decision or an input or they're done with the process or have detected something.

The human's going to take it down to whatever interface it's going to-- VS Code, CLI, some sort of kind of chat interface, whatever they're most comfortable, whatever is most appropriate for the task and then kick it back up and that's going to be the world of every single knowledge worker.

Dori: All right, one last question before the lightning round.

Scott: Okay.

Dori: What's the most painful bug or failure that you've seen in production and what did you learn from it?

Scott: What I found with data is-- There was a comedian that once said like when an escalator breaks it just becomes stairs. I mean it's different but it's still useful. Like when your data workload fails, it just often results in stale data. And I think that's like the thing that I've learned is it's kind of just manage expectations, I guess. In the data space. Yes. There are some things that can go wrong.

And don't get me wrong, we usually have tests and things like that, but understanding the difference between a real error and hey, the pipeline didn't run, so the data is, you know, three hours old instead of one hour old. And I think it's just understanding the criticality, the severity of an issue is what kind of taught me.

Because again, I used to early in my career, panic, you know, okay, that didn't run. But what's the real impact? The data's two hours old? I think we'll be okay.

Dori: Yeah.

CL: Okay, so separating the failure mode for different levels of expectation.

Scott: As always, you summarize it much better than I said. Exactly. I totally agree. What do you say?

CL: All right, Scott, this has been so fun. And we'll get into the lightning rounds. Just real quick questions and short answer. Are you ready?

Scott: Sure.

CL: Okay. First programming language you loved or hated.

Scott: I'm sticking to my guns. Excel.

Dori: Full circle.

Scott: Yeah, I'll be the first one that said that.

CL: All right, your go to data set for testing a new tool or a new methodology or whatever?

Scott: Ooh, that's. I mean, you know, I like just some census data or something like that. Something pretty accessible and understandable.

CL: Okay, what's one lesson from outside of tech that influenced how you build?

Scott: My early days in consulting really taught me, I guess, the importance of a job well done, of quality, and the pursuit of excellence. And I think we were helping our clients try to make real important decisions. And I think just the hunting for a really great answer and the commitment to doing great quality work is a lesson that has always stayed with me.

Dori: Yeah. What is your favorite podcast or book that's not about data?

Scott: What's interesting is I actually don't listen to a lot of podcasts about data. Like, I feel like I've spent so much time in the data space. You know, I've always listened to the Economist, you know, the audio version of every issue. I feel like great source of information on what's going on in the world, which even in your day to day as a data person is tremendously helpful.

Dori: Yeah, no arguments for me there. And then final, final question. Tabs or spaces?

Scott: It kind of depends on the language. But, you know, I would say I'll say tabs.

Dori: I'm also a tabs person. We have not had many tabs people on the show.

Scott: I thought you were, Dori.

Dori: Yeah, the shortcut. Awesome.

CL: Where can listeners find you and be helpful to you?

Scott: Yeah, I mean, find me on LinkedIn. I mean, I'm one of those cool kids that, that's my primary social network. So, like, I'm on Twitter as well. But, you know, just search me on LinkedIn, add me or I'm Scott@kilocode.AI.

Dori: Well, thank you so much, Scott. This has been an absolute pleasure.

Scott: This has been a lot of fun.