If you rely on complex scaffolding to build AI agents you aren't scaling you are coping. Thibault Sottiaux from OpenAI’s Codex team joins us to explain why they are ruthlessly removing the harness to solve for true agentic autonomy. We discuss the bitter lesson of vertical integration, why scalable primitives beat clever tricks, and how the rise of the super bus factor is reshaping engineering careers.
Show Notes
- OpenAI Codex: Learn more about the models powering tools like GitHub Copilot.
- Codex Open Source Repo: The lightweight coding agent that runs in your terminal (check out the Rust migration mentioned in the episode).
- Agent Skills Open Standard: The open standard and catalog for giving agents new capabilities.
- The Bitter Lesson: Richard Sutton’s essay on why compute-centric methods win in AI.
- Follow Tibo on X @thsottiaux | GitHub
Transcript
(Disclaimer: may contain unintentionally confusing, inaccurate and/or amusing transcription errors)
[00:00:00] Andrew Zigler: My guest today is Tibo Sottiaux from OpenAI's Codex. Team. Tibo has been working shoulder to shoulder with research and engineering in San Francisco to solve one of the hardest problems in tech, true agentic autonomy. Tibo, it's really great to have you here on Dev Interrupted.
[00:00:15] Thibault Sottiaux: Hey, awesome to be here. Uh, really excited to talk about all of this.
[00:00:19] Andrew Zigler: Great. So today we're gonna look at the bitter lesson of AI development. You know, there's so much we can dig into about what you're building and about why clever tricks and domain expertise can sometimes fall short, and why scalable primitives are winning. We're gonna explore the Codex team and how it balances that research with exacting the requirements for production and making a happy developer tool that works for everybody.
[00:00:43] Andrew Zigler: Um, but I wanna start at Codex, just in your own words, like learning about OpenAI's flagship coding agent. You know, can you tell us a bit about what it is and why you describe it as like an agent first instead of a product?
[00:00:56] Thibault Sottiaux: Yeah, so we think about the two parts, like first and foremost we're [00:01:00] building a SOTA agent that is able to act and perform, uh, incredible amounts of like, work, uh, on, you know, the coding front and help software engineers in this world. this agent is fairly general. You can put it to work in many places, and that's also where the products come in.
[00:01:16] Thibault Sottiaux: So it's like figuring out. What the best way is to leverage an interface with this agent that is going to be ever increasingly more capable. Um. And there is something really interesting when you shift your mindset to building an agent first and then figuring out where to put it to work. It's like you, you find like a remarkable amount of places where this agent, you know, comes in handy and, you know, can actually do economically valuable, uh, work. And we're also thinking about, you know, what does this mean beyond just coding, right? Like even for software engineers, uh, it's not just about code generation, it's about solving many other parts of like the day to day that are actually the bottlenecks. Uh, and so. You know, building that general agent is like what we're after.
[00:01:55] Andrew Zigler: And, you know, we're gonna talk more about what those other bottlenecks are. We [00:02:00] talk about that a lot on Dev Interrupted, especially like in the last year, about how, um, agentic development has really made bear all of those human and communication problems that are, you know, affecting teams really deeply.
[00:02:11] Andrew Zigler: Something that you said there about, um. Thinking about it differently from being a product, it actually makes me think of almost like a product and a platform analog. With a platform you can put a lot of products, a lot of things in an ecosystem together. An agent, it seems to be emerging as something that operates the same way.
[00:02:27] Andrew Zigler: You can build an agent, like you said, and then figure out what are the use cases, how do we step this forward or backward into the products that make sense for people? Yeah.
[00:02:36] Thibault Sottiaux: right. So if you think about, you know, building an autonomous, like ever more capable entity, um, and keep your mind sort of flexible about, you know, what the best form factor is, what the best product around it is. And, you know, perhaps we put it to work inside our own products, but we also, you know, partner with other companies to put it to work in their products.
[00:02:55] Thibault Sottiaux: Uh, it opens the door for like a lot of like, you know, great ideas that you don't necessarily have, you [00:03:00] know, we don't have them ahead of time, right? It's like, you know, we
[00:03:02] Andrew Zigler: Right.
[00:03:03] Thibault Sottiaux: on like building the agent and then, you know, figuring out like where to put it to work later.
[00:03:07] Andrew Zigler: And, and what is it like to work in an environment where you're building a coding agent and you're sitting on top of a frontier model? Like you have the vertical integration of that entire process, which is a unique leg up. What is that like?
[00:03:20] Thibault Sottiaux: That's right. It, um, it's very interesting because we're able to take some of the best ideas from engineering and have them influence research and then do the reverse as well, where like research influences like the entire, like engineering roadmap for, you know, how we build out the agent. And you know, one of the things that you can do, like when you vertically integrated is like you can decide where you actually fix problems.
[00:03:42] Thibault Sottiaux: So you don't have to fix everything in your harness. Some of the things we decide to fix, you know, downstream, like by training new models, um, and we know that, you know, by training the model, like we will have a jump in the capability that we need, like, you know, three months down the line, six months down the line, and allows us to do these trade-offs, um, that you can't do without like a [00:04:00] vertical integration. There's also this thing called, you know, the no free launch tier, which is basically like if you're trying to adapt and be intelligent in any possible distribution, um, well, this is going to be strictly less optimal that if you were to actually something for a very specific distribution. And so by coupling like the harness and the model and like that's what we named the agent, we're able to get a lift in capabilities.
[00:04:26] Thibault Sottiaux: And that's, you know, obviously very interesting. Like, you know, for something as important as coding.
[00:04:29] Andrew Zigler: Yeah. And it's fascinating that you'd mentioned too that those two things, they feed into each other. The research and the engineering, it's a big loop and inside of it, there's lots of little loops. I think that's what we're, we've all been discovering in the last year, especially people who have a lot of success working with agents, is finding those loops, those workflows that work perfectly, the inputs and outputs, uh, feeding into each other and making the system.
[00:04:52] Andrew Zigler: Better. Um, so that is a unique advantage and kinda like a vertical integration front. And I guess in this world [00:05:00] where there's so much greenfield, you have the model, you have the coding agent, you can build it and take it any direction. Take this almost like, uh, agentic first approach and, and then build towards what individual people and organizations need.
[00:05:13] Andrew Zigler: But how do you, as the person at the center of that prevent complexity? From trapping your own architecture and keeping it simple, that way it works for everybody.
[00:05:23] Thibault Sottiaux: Yeah, we, so we try to stay grounded in things that, you know, have stood like the test of time and that we know, you know, have been true in the past. Um, where simplicity just often leads to better scaling over time and better overall performance of the system that you're building over time. What we're really looking at is the ability of, you know, the agents.
[00:05:45] Thibault Sottiaux: So it is like this, the system of like the harness and the model to continue to increase in capability over time as our models continue to increase in capabilities. So it would be, would be really unfortunate if we [00:06:00] introduced bias in the harness or in our solutions, you know, which means that when we have a capability jump into model, like we're actually not able to express that, you know, because we somehow constrained it, um, in a way that prevents it from expressing its true capability. And, you know, that's what we would refer to as like a capability overhang. And by keeping things simple, like you reduce the amount of things that can actually go wrong. Um, and it's all about choosing the primitives that are proving to scale with model capabilities. And, you know, that's actually a also like closely really true like a research problem.
[00:06:35] Thibault Sottiaux: Like we can do the research on that of like, hey, it's like this harness scales with, you know, different model sizes and different model capabilities. Like, you know, we look at. It's like on like the smallest model and then like a mid model and then like a frontier model. Like, you know, how do we do, we get the scaling that we expect, so you can extend sort of like the scaling laws to, um, you know, scaling laws the entire system.
[00:06:59] Andrew Zigler: [00:07:00] Right. And, and so in order to apply the scaling laws, you have to have simplicity ultimately, that those, those, those things are leveraging, that they're driving. Otherwise things are gonna clash. They're not going to work. And like you said, you can't take advantage of immediate gains. I think that's really the biggest, you know, win of this system is that when you're so lightweight, it's so easy to pick up and put it onto the next new thing, which is what's so important when you're coupled closely with the model.
[00:07:25] Andrew Zigler: And, and so that's, that's something that's really fascinating. I think the, the simplicity though, you know, it can't be understated, it's earned, right? It's like earned through research and engineering. Like you had the climb a summit of complexity, I'm sure. And at one point things were very complex, and then you have to come down from that summit, like you said, the find the primitives, the things that emerged as more clear.
[00:07:49] Andrew Zigler: Right. And, and, and so that's like the journey that, that you're kind of describing.
[00:07:53] Thibault Sottiaux: Yeah, it's, it's an.
[00:07:54] Andrew Zigler: I,
[00:07:55] Thibault Sottiaux: It's an ever continuing, like search for the right primitives. And you know, once you [00:08:00] discover those primitives, they seem delightfully simple. Uh, they often seem delightfully simple,
[00:08:05] Andrew Zigler: yeah.
[00:08:05] Thibault Sottiaux: for those primitives is, um, it's a, it's a complex matter and you might not find those primitives immediately.
[00:08:12] Thibault Sottiaux: And so, you know, you have like, complexity in your system and you, you look at that complexity and you go, huh. Um, we think like there's quite a bit of bias here that we introduce and, you know, we constrain the model in a way where, you know, it will hamper its, um, you know, productivity at some point. Um, but this is also like this delicate balance between the harness.
[00:08:31] Thibault Sottiaux: Um, and, you know, to me it's like called a harness because You're scaffolding it, uh, in a way where you want to like, remove the scaffold over time
[00:08:41] Andrew Zigler: Right.
[00:08:41] Thibault Sottiaux: is able to stand on its own. you know, it's all about figuring out, like, you know, when is the right time to remove like, pieces of the scaffold, uh, and to build in a way where you can do that.
[00:08:50] Thibault Sottiaux: And the advantage of like coupling the two together is like, you know, we have to care about like, our model series and our model series only. And so every time, you know, we improve things, we can [00:09:00] remove bits of the scaffold without, you know, being afraid of like breaking something that is not under our control.
[00:09:05] Andrew Zigler: That's a, that's a really smart insight that I wanna like click on again, because you know, you're calling out that ultimately the harness, that scaffold that you build, it should lean towards that simplicity because like you said, it's supposed to be kind of there until the agent can stand fully on its own and do what that harness is kind of getting it to do.
[00:09:22] Andrew Zigler: Some people lean, of course, the opposite approach and they treat that harness more like a jet pack and it doesn't go the way of simplicity. It goes of how can I put so many tools and so many things in this box and, and so I think it's like two different types of ways of people come at them. Um, and they obviously achieve different levels of success.
[00:09:41] Andrew Zigler: It doesn't scale as well as having a lightweight harness, like what you're calling out. And I, I think that's kind of part of what y'all have, earned from your years of research. And, you know, there's also, um, Richard Sutton's, uh, bitter Lesson, which calls out that clever tricks and domain expertise, they don't scale compared to these [00:10:00] primitives.
[00:10:00] Andrew Zigler: Right? And when we talk about these primitives, they're compute and power and time. Uh, and that's what you're on the mission to find is how can we take best advantage of, of those so everybody can get it. But, um, you know, along that journey you mentioned that you open sourced like the repo, you open sourced the process by which people can build these agents.
[00:10:19] Andrew Zigler: What was that like? Like what led y'all to that decision?
[00:10:22] Thibault Sottiaux: So around the time that we decided to open source of repo, I think there was a lot of. Um, mysticism around, you know, agents and you know how they worked and you know, what you needed to build on top of a model in order to have it, you know, be able to act safely into the world and like perform. Uh, economical valuable tasks for you. So we wanted to sort of show how, you know, delightfully simple, it can actually be and show some of the primitives that are important to get right. And, you know, we also wanted to show that if you do these things, you know, you can get incredible performance out of our models. And so, [00:11:00] you know, it plays this like dual role.
[00:11:01] Thibault Sottiaux: Additionally on top of that, sort of like had this idea that know if you solve for code generation, open source is going to change. And we want to deeply understand how open source is going to change by being open source ourselves, at least for part of our technology. And then one additional, idea and like reason for why, you know, I thought this might be interesting and like others agreed, was that we were going to see a lot of tinkering in the space of agents and a lot of interesting products being built on top of it and having like an open source core, you know, allows people out there in the world to like, you know, get inspired and tinker and like, build things that you, we might have never, you know, we might have never imagined. Um, and so it allows like this sort of like, to, to tap into this creativity of like the open source community and to let people like, you know, innovate and invent new things. And like, I think altogether it just made so much sense, um, for the agent to be open source and, uh. Yeah, it's like it is something that I've been very [00:12:00] proud of.
[00:12:00] Andrew Zigler: Yeah. And, and as part of like building that community, making a place where people can come and be part of, uh, the development, have there been things about that that like surprised you or stood out where you're like, I'm so that, like this was an earned thing from the community. Like this is a proven success of why open sourcing.
[00:12:16] Andrew Zigler: This was the way.
[00:12:17] Thibault Sottiaux: yeah. One of the things is initially we got things somewhat wrong, like we accepted like too many contributions and you know, sort of like lost control over the repo. Uh, we course corrected later, like at some point we migrated to rust, uh, and we had decided that, you know, we had conviction on like some of the primitives that we needed to build and, you know, we wanted to build those primitives.
[00:12:39] Thibault Sottiaux: Right. And, you know, we also had conviction dot. We were going to see, you know, millions if not billions of agents running concurrently at some point. And we, we wanted to write this like in an official language. So like the rust migration was a bit of a, a difficult moment with the community 'cause we had been accepting a lot of prs through like our types script, um, [00:13:00] open source repo at the time.
[00:13:01] Thibault Sottiaux: And then, you know, we just sort of like rewrote things to rust. We did build like very good, Partnerships with like awesome contributors over time to the rust core. so that, that relationship to the community has evolved over time. Uh, one of the things is just how great it is to have it all there out in the open is like, we do get a lot of bug reports and people like sort of like point at, you know, this is where it is.
[00:13:25] Thibault Sottiaux: Like, you know, here's like how you should fix it or like, you know, here's like an attempt at fixing it. And then we also get a lot of inspiration from all the forks. Um, I think there's like. don't know the number, but like over a thousand forks of the repo. Some of them are actually quite popular. Um, and, you know, they just come up with like, new ideas and like we, we, we work with the authors of like the, the forks, like, you know, to port some of the changes back into, um, you know, the, the, the, the Codex Open search repo.
[00:13:50] Andrew Zigler: And it is pretty cool, like to be able to learn those things from how people change and alter your code. And you can also then get a glimpse of how the agent can be adapted to other things. You're gonna [00:14:00] get a a, a, a really good front row seat at how these early adopter developers are picking up these systems and molding them to their specific domain problems.
[00:14:09] Andrew Zigler: So it's a really great partnership and um, definitely something I think people should. Should check out because that's, that's kind of rare and it's part of that relationship that even goes back as deep as the model, right. Um, that's kind of how deeply penetrating this type of technology can be.
[00:14:23] Thibault Sottiaux: every, every week I hear from a company that, um, you know, builds their company around like the Codex Open source agent. where we're like, oh, you know, it's just like, it works so well. It works actually, like on non-coding things. And we adopted it like, you know, we adapted it like in this, in this way and like now it's able to do like this other economically viable task, um, such as like editing spreadsheets, for example, and, or like we embedded it into a browser.
[00:14:49] Thibault Sottiaux: Um, and then I get these demos and like, it's always very cool to see that, um, you know, somehow we're like contributing to those things as well.
[00:14:55] Andrew Zigler: Yeah, totally. I, I'm curious too, to know, like, you know, you talked about [00:15:00] like earning all of this simplicity and getting back to the primitives and, and you know, we're at the top of a new year. We spent all last year experimenting with these new tools. You know, what is something that you picked up in 2025 that maybe you're not gonna pick up again in 2026?
[00:15:14] Andrew Zigler: Like, how is your workflow in the way that you think about agents getting simpler this year?
[00:15:18] Thibault Sottiaux: Yeah, so one of the big issues and challenges that we had last year was like very long running sessions, um, where the agent goes through something called a compaction.
[00:15:29] Andrew Zigler: Mm-hmm.
[00:15:30] Thibault Sottiaux: Where, you know, we allowed the agent to perform work beyond the context window that is actually available to the model. And a lot of the complaints were, you know, this is not working well.
[00:15:40] Thibault Sottiaux: Um, and this was like really part of the scaffolding at the time where, it was just sort of like a heuristic of like how you needed to summarize the work done so far and then, you know, like reset the context window so that the work could be allowed to continue. Um, you know, when given a fresh, um, and. The, the model was like [00:16:00] losing context on like, quite a bit of the work that was performed before. Um, it's really hard to get this heuristic right. Um, you know, you can try and prompt models. You can try and like, you know, scaffold your way through it. Um, for a lot of. agents out there. This is like a very significant part of like the complexity of the harness. And so we decided to solve this like, you know, at the model level and, you know, we train on this like end to end and now this is something that we receive like almost no complaints on anymore.
[00:16:26] Thibault Sottiaux: And, uh, you know, the. model is able, like, the agent's able to like work across like 20 windows, 20 context windows, like, you know, for very, very long, um, time horizons without losing track of like, you know, what it was doing before. So that's like something that we're, we just, we remove the whole bunch of complexity just by solving it, you know, in a way that we're, we're uniquely positioned to solve it.
[00:16:47] Andrew Zigler: and that's the, that's the great thing about like the, the kind of trifecta, the simplicities of those simple primitives is when one moves and gets a major advantage. When you come back to another one, you realize that you can move it way further because of all of these gains, [00:17:00] they're so closely coupled.
[00:17:01] Andrew Zigler: Are there any other, ways that you're approaching Codex itself for like this year that stand out for that, like, will be different you think from, from how you approached it next year? Like what's like the top of the year vision for how we make this like the most popular code agent in the world?
[00:17:18] Thibault Sottiaux: Yeah, so there is a couple of things that we're really pushing on. We think last year, agents became reliable enough that they can perform work. Like this year we're going to see like reliable multi-agent networks, uh, performing significantly more work. What does that mean for the user when you have, you know, maybe an order of magnitude or like two orders of magnitude more, you know, economical valuable work being produced in, you know, the same time span. You know, it means that you're able to spend like a lot more tokens in the same, um, in the same time, but like also means like perhaps you need to review like a ton more code. Um, or like, how do you do that exactly? This is something that we're looking to solve. We're [00:18:00] working towards, you know, making things significantly faster as well. We feel like we are at the frontier with our models on the intelligence front. Like we're not, you know, at the frontier, um, yet on like, you know, how fast it can be. Like we expect these models to get like, significantly faster this year and it's almost like sweet spot, uh, that you want to reach of like a level of intelligence that, like sufficient speed that, you know, it feels delightful to use in the product. Um, so that's one thing that we're working towards as well. And like, and. The last one I would say is like, you know, it's really this super collaborative personality. Um, our models, like in our agent like Codex is known to be a little bit terse, um, a little bit stubborn sometimes, you know, just really this blunt, pragmatic engineer, um, that is not the right persona for everyone.
[00:18:47] Thibault Sottiaux: Um, personally, I, I like, like, you know, to be a little bit more validated in my ideas. Um, you know, not to be told what, that I'm right when I'm not right. I want the model to acknowledge that I'm there, you know, behind the laptop as well, and I trying to do [00:19:00] something and to
[00:19:00] Andrew Zigler: Totally.
[00:19:01] Thibault Sottiaux: the way with me.
[00:19:02] Thibault Sottiaux: And so, you know, solving for that is important as well.
[00:19:04] Andrew Zigler: I know, I, I know what you mean. I think that's like something we've experienced with all of the emerge, like front foundation models is they all have like a different level of politeness, terseness, and like you said, like sometimes it's too much. Politeness is just a waste of everyone's time. It's not accurate either.
[00:19:16] Andrew Zigler: So, um, but there's like a happy medium there. Like you said, you can't, like what personality works for everybody. No personality who's liked by everybody. No one. It's like, so how can you build an agent that built by. It has liked by everybody. You can't, but you have to make it easy to adapt to what people do.
[00:19:33] Andrew Zigler: Like.
[00:19:34] Thibault Sottiaux: That's
[00:19:34] Thibault Sottiaux: Yeah.
[00:19:34] Thibault Sottiaux: it's really about making it personal to you. Um, and, you know, working in the ways that you know, you like, uh, working. So if you are like highly creative and you like brainstorming and you know, you don't want to like get like. Pedantic, like knits on like the quality of your code. It's like that should be possible.
[00:19:53] Andrew Zigler: Right.
[00:19:53] Thibault Sottiaux: you're working on like a super critical code base and you know, you want every single thing that can go wrong to be flagged to you, [00:20:00] it's like, you know, then you know that you should also be able to get that. And that's like something that our models are incredibly good at. Um, Codex is being, you know, involved in finding some of the most, um, impressive exploits last year, you know, some of the react exploits, um, which, you know, we can talk about, You know, that you don't need like a super friendly personality. You just need Codex to go and, you know, figure out like, you know, what a potential exploit is or, you know, like fix like a very gnarly bug for you and then come back with a solution. And you want to be very confident that it's right. Um, you know, for others it's like, you know, you just really want this bubbly like, you know, super collaborative thing.
[00:20:37] Thibault Sottiaux: Um, you know, and it needs to be personal.
[00:20:41] Andrew Zigler: so speaking of like being able to customize and couple things together, like you are in an environment where you have the distinct privilege of being, like we've talked about, connected to the foundation model all the way up to the implementation of the agent and ev all the research in between.
[00:20:55] Andrew Zigler: So the full stack. But for other teams that are building with [00:21:00] agentic tools and are, and are and are and are making these kinds of age agentic systems, like how should they think about coupling, like really tightly coupling to a foundation model versus being more portable in what they build. And I'm curious like how that affects the simplicity or the primitives that we've been talking about.
[00:21:20] Thibault Sottiaux: Yes, I think the primitives, um, should roughly be similar, although they might differ in shape a little bit. Um, and I think one of the challenges is like if you're building for, for many foundation models and you know, are completely agnostic to the model provider, is you have to find a common ground of all these models. Um, and it's, it's quite inevitable that if you do not adjust at least a little bit, um, you are going to see, the penalty and like, you know, the overall performance that you can get. And so what we're seeing is like, you know, we're working with, you know, some, some of the players in the ecosystem and like [00:22:00] we're advising them on like how to make it work very well.
[00:22:02] Thibault Sottiaux: And this is also why our code is open source. It's like it's just all out there. It's an example of how you can get the best performance, you know, out of the GPT models and, you know, we work with them in order to adjust things. So I do think that you need to adjust it to like a certain degree in order to benefit from the models. At some point if you want to do that for, you know, the dozens of models out there, it's like, it becomes like quite prohibitive. Um, and I expect like, you know, major players to like, you know, only do that for like a handful of models.
[00:22:31] Andrew Zigler: So we've been talking about the agent, but I wanted like us to take a step back and talk about the developer, the engineer, the person sitting at the computer using the agents and, and, and what their new normal looks like. And that's what we've all been exploring in the last year of how this type of agentic coding is impacting developer teams and individual engineers.
[00:22:51] Andrew Zigler: Different people are getting at this at different speeds, right. And with it, you know, the folks who are experiencing that hyper productivity, it, it really [00:23:00] blows away all of the code bottlenecks and new co bottlenecks emerge. And they're often around like planning and integration, actually taking the time to like sit down and chat with somebody.
[00:23:09] Andrew Zigler: So, you know, there's a new skillset. Involved with what makes a developer successful. In a way, each developer almost becomes like their own mini team that they can customize in whatever way works for them specifically, and they become the representative, almost like the decision manager of this team. You know?
[00:23:28] Andrew Zigler: What has that, what has that been like for your engineering team? How does their day-to-day work transform in that kind of new world. Um, do, does everyone just spend less time in an IDE and just lots and lots of time planning and reviewing?
[00:23:43] Thibault Sottiaux: Yeah, so interestingly, it actually has brought, um, people closer together and, you know, we have like more FaceTime, we have more, you know, creative like, uh, ideation and planning together because everyone is so [00:24:00] accelerated that once you decide on the thing and you can just like almost immediately do it. And agreeing together and, um, getting organized together, you know, as like a small team, even like of 20 people. Um, it's like, it matters a lot because you'll be able to achieve like in a week what you traditionally were able to achieve, like, you know, maybe in a month with the team of like, you know, the same size. And so it's kind of like interesting to see that it does bring engineers, you know, to like talk more and align more maybe like, uh, upfront than, than before. We're also thinking about, you know, obviously like, hey, how can you help and accelerate, you know, the planning phase and like, you know, building consensus and like, you know, routing things into like, you know, user feedback and the realities of like, you know, finding PMF in these things. Um, one of the bottlenecks that is like downstream of that, you know, is code review as well. And this is something that we proactively thought about like last year we built a custom like code review model, uh, which is SOTA, I [00:25:00] believe, like it's still SOTA, um, today. And we deployed it like internally across all of OpenAI.
[00:25:06] Thibault Sottiaux: And this has been. To my surprise, one of the largest successes, like within OpenAI, you know, for Codex where it's not like pretty much enabled for, for everyone by default. Um, a lot of teams just like require, it's like mandatory to have Codex like reviewed the PRs because it catches so many bugs. Um, and it's inevitable that, you know, if you're generating so much more good, it's like, you know, you're also going to produce like, you know, some of, some amount of good that you then, you know, want to catch. Figuring out like what those bottlenecks are starting to become and like solving for one, you know, for them, like one at a time. Um, I think that's like the big challenge of like 2026.
[00:25:44] Andrew Zigler: Yeah, I think so too. And, and those bottlenecks, those are things that are popping up. Those like problems like groundhogs, like popping outta the ground, right? Like for these teams to, to solve because they've already been in motion with each other. They're already working. But there's also like this, a whole other wave [00:26:00] of people right now who are finishing up their CS degrees who have been experimenting with these tools while they've been studying, who are, you know, getting their first jobs as, as engineers.
[00:26:08] Andrew Zigler: And they're not coming in with all of that baggage. They're not. And so they're coming in very, very differently. And you. From my own personal experience, like dealing with, with, uh, like folks that are in that skill band, like they're absolutely crushing it with these tools. When they use it, when they lean into it, they have like a, a native instinct because they aren't cluttered and bogged down with all this stuff like you and I are from just being in like the working world.
[00:26:31] Andrew Zigler: Uh, so like what, what do you see, like with OpenAI, I know y'all work with a lot of young developers and folks that are, that are engineers. Like what is that difference been like in how these like agent first developers are coming on the scene and um, how do they solve things?
[00:26:46] Thibault Sottiaux: Codex we have the entire spectrum. We have like industry veterans who have been like, you know, working, you know, for over like 35 years. Um, I
[00:26:54] Andrew Zigler: Wow.
[00:26:55] Thibault Sottiaux: the record. Um, you know, it's like before, uh, before I was born. [00:27:00] Um, and like, and then we all, we have new grads, um, and one of the. One of the people, like I trust the more, uh, trust the most, like on, on the team is like, you know, this new grad, um, Ahmed, he's just got this like incredibly creative ideas. Uh, almost was like, you know, molded by, you know, all of this, like existing, like, has never really built these habits, you know, and like decades of like, you know, software engineer of like, oh, this is how you do things.
[00:27:25] Thibault Sottiaux: Um, it's like super open to new ways. Very, very much adapting every day, um, and teaching a lot of the rest of the team, like, you know, how to actually be productive. And it's like remarkable to see and like, you know, we have a couple of others, like that on the team as well. And it's essential, uh, for us, I would say, like, you know, without, that is like, we would actually be moving like way slower as a team.
[00:27:49] Andrew Zigler: Yeah, I agree. I think, I think you start with those, like there's individual movers within your org and I think every org has an ident, has a responsibility to identify them. They're there, they're doing something. If you [00:28:00] don't know about them, I promise you they're lurking somewhere under the surface. And if you find them, there's so much that they can teach and enable to the rest of your org because so often during like AI rollouts and people experiment with these tools, things get like trapped inside of teams and they don't really get broader than that.
[00:28:15] Andrew Zigler: Or you get one high powered individual who is able to just massively ship everything and no one else really knows how, any of that how, how that guy does anything. Right. And so, you know, it's, and, and speaking of it kind of makes me think of the emerging like super bus factor problem, right? The idea that like if, if a single engineer, like you said of the, the biggest bottleneck is that's just a siding and then somebody fires off an agent and then it gets done.
[00:28:39] Andrew Zigler: If the big, if you know, a single engineer can ship a whole product solo that we're seeing what people have to be able to do with the kind of tool, like how do you keep collaboration alive? Uh, in that kind of world, why, why even bother handing anything else to someone? Me and my army of agents that I can spin up on demand can just do it.
[00:28:57] Thibault Sottiaux: Yes, maybe in the limit, you know, [00:29:00] you'll be able to like run a company solo. Um, and it's like very much like, you know, maybe like, you know, five, 10 years away, uh, in my head, like maybe we get there sooner. And, you know, I think that will be like an interesting time, but like right now what I'm seeing is that disseminating plans is important.
[00:29:19] Thibault Sottiaux: So disseminating information and like intent, and recording intent of changes is important. you know, an agent can help you, uh, achieve that as well. So like, you know, as you go through your conversation and a session, you know, you can sort of like get a summary of your intent, attach that to the PR later so that, you know, others can like, understand. Um, I've started to build tools, you know, for myself and like, uh, for folks in the team as well like, you know, to keep track with changes that are happening, you know, across our team, across the organization, you know, at different levels of abstractions, um, because there's just like, there's a, there's more happening. Um, and you want to offset that by, you know, also allowing to sort of like, give everyone a superpower to [00:30:00] understand things faster. Um, understand, you know, what is changing, understanding how things are implemented, why things are implemented in a certain way. So it's not just about, you know, making co-generation like a hundred times faster.
[00:30:12] Thibault Sottiaux: It's, it's really about, Giving a boost to like how quickly you can, you know, find the right ideas to solve for and how quickly you can understand as well, like, you know, as a human, like the state of things. Um, so that's like very much like top of mind for us is, you know, solving towards that.
[00:30:29] Andrew Zigler: You mentioned this world, like getting plans out there so people can see them and understand. And I think that's really important. And I, I, I've seen a lot of tools in a lot of companies try to tackle this, right? Like everyone has taken a bite out of the whole spec problem, making a plan. And I'm, I'm curious that solves a lot of things that, like, it addresses the human to agent handoff, but also agent to agent. Agent to human. Like what, what do you think about that as like a, as a primitive, just like the rest of these things that, like it's, if you, if it has to be in there, then how do we make it as simple as [00:31:00] possible?
[00:31:00] Thibault Sottiaux: Yeah, so the, downside of like relying only on plans and, you know, building a large spec over time of your product or implementation is like, you know, it gets like a little bit, you know, too wild. And then, you know, you find like contradictions in there as well and you know, at some point it's like so big that's like, it becomes like inscrutable as well. Um, and maybe there is a mismatch between the plan and like the implementation, but. I do, I am a big believer in, you know, even, you know, before all of this, like, you know, things like design docs and, you know, getting, you know, your ideas together for like, you know, where the product is going and, you know, writing down your vision, your strategy.
[00:31:37] Thibault Sottiaux: And so like, I think that is like ever more important. Um, at the same time, iterating on new ideas has never been, you know, easier. And, you know, gaining like that signal and that information that you need in order to make a product decision, you know, has never been easier. This is like really something that's greatly accelerating. And so sometimes you might not know what you [00:32:00] actually need to do, but you know, like kinds of things that you need to build in order to gain the signal that you need, uh, in order to know what you need to build. Um, and so, you know, sometimes the plan is just like, we need to gain the signal. You know,
[00:32:12] Andrew Zigler: Right.
[00:32:13] Thibault Sottiaux: things that we're gonna do.
[00:32:14] Andrew Zigler: Yeah, we need to build the lightning rod so that the lightning strikes it so we can figure out
[00:32:18] Thibault Sottiaux: Exactly.
[00:32:19] Andrew Zigler: on. Right. And that and that, and that is what's free now. Right. It's like being able to just like, you know, you can spin that. Like you said, it's easier than ever to investigate something about your product, how people are using it, like all these different types of things than the meta space around your product.
[00:32:35] Andrew Zigler: Right. Things I can read now solve with engineering that before would've taken like deep data analysis and a lot of time, and a really like clear thesis upfront. Now you can just like freeform ex explore, like sometimes even in parallel with like different ideas, right? So, uh, like if that is the new
[00:32:53] Andrew Zigler: paradigm by which developers build. Then what do you think the career path looks like for like senior IC or becoming like a [00:33:00] staff engineer is like that seniority defined by how well you can orchestrate those agents and, and your plans and, and then I guess share it with your colleagues at at scale.
[00:33:11] Thibault Sottiaux: I don't think it's fundamentally different, um, to, before I think like, you know, the path through like, you know, senior staff and beyond, it's like, you know, the impact that you have like within the team, you know, within your org. this is, you know, how I, I've always. Thought about it is, you know, like how effective are you, you know, at like building like impactful pieces of the product. and then, you know, making that work in the organization, making that work for the company. And so there is an aspect there of like, yes, it is increasingly like you have to scale yourself up. You can do so many more things like solo you can go in like look at user feedback. You can look at logs, you can, you know, run a few queries in the background, you know, so that you understand like, you know, which database schema, you know, like is the most appropriate.
[00:33:57] Thibault Sottiaux: You know, so that like the queries that you have in [00:34:00] production are actually going to be like running performantly and you know, you can run like, sort of like a whole like little engineering team, you know, by yourself. And so like the skills, you know, that are interesting and that you need to build, like really as an engineer is like more and more towards like everyone, you know, growing to like a tech lead role or like a tech lead manager role, um, together with like, you know, wearing like, you know, an increasing, um, increasingly large like product hat as well.
[00:34:25] Thibault Sottiaux: Like, you know, building that empathy for the user and, you know, models and like the agent can like, help you with that as well, like over time, because it should, you know, you should also be able to send like the, the agent, you know, like, interview users or, you know, summarize, you know, what the internet thinks about your product and like, you know, potential things that you should try. Um, and yeah, so just like maybe this accelerating path towards like, you know, TLM type roles, uh, I would say a lot of, a lot of people, you know, that's what they've always strived, you know, to do. So I
[00:34:55] Andrew Zigler: Yeah.
[00:34:56] Thibault Sottiaux: compatible.
[00:34:57] Andrew Zigler: No, I think that, I think that's great advice and it kind of, [00:35:00] it mirrors with a lot of what we've been hearing, especially on the show in the last year, what people have been talking about, the skills that matter most of them now. And I just, the end, um, our, our conversation and we've covered a lot of amazing ground.
[00:35:11] Andrew Zigler: We got a really great look at how you think about Codex and how you're tackling the coding agent problem. But from your perspective, you know, you have an incredible vantage. Um, any final advice? You want to impart to our listeners how they can future proof themselves or how they work or at the top of a new year, so I'm sure there's lots of new ideas we could tap into.
[00:35:31] Thibault Sottiaux: Yeah. So one, one advice is like, one thing I've had a lot of fun with, um, and a lot of people in my team at OpenAI that fun with is like skills. Um, so this is like now an open standard and Uh, it's like a little thing that you can teach the model to do. like in the way that, you know, you think is like, you know, most effective.
[00:35:50] Thibault Sottiaux: You know, for example, like looking at logs or like, you know, running a performance test or like, I have like this QA skill, like where Codex is able to like QA itself. Um, so like [00:36:00] whenever I build a new feature and, you know, I just send Codex like, you know, to play with a version of itself in the terminal, like make sure that it's implemented up to spec. Uh, and there's like no regression, like building, like finding skills and like really my advice is like to make them your own so that you know you're building the skills that you need from your agent over time so that they're adapted to your workflow. Um, and I, my analogy to is like, you know. The other day I was like, thinking about it.
[00:36:25] Thibault Sottiaux: I was like, this is the closest I feel to like having trained like a little Pokemon, uh, you know, where I'm like, oh, you know, this thing is like leveling up. You know? Every
[00:36:33] Andrew Zigler: Right.
[00:36:33] Thibault Sottiaux: with it, I'm like, learn this new thing. And it's just like, oh my God. It's like doing it like a little bit better every time now. Um, and that is like, you know, it starts to feel like this sort of like reliable, like, you know, you're building this, this, almost this bond, um, because it's like more and more reliable over time and, you know, your work gets more delightful as well because you're automating the parts that you know you actually want to automate and that you know that you don't want to do. There is this pitfall as well, [00:37:00] you know, which this is why I'm recommending this, like, this pitfall of like, you know, just only automating code generation, but like, if you use skills and think about and like all the other things that you wanna automate, it's like actually you can keep like, you know, the most delightful parts, you know, of your day.
[00:37:12] Thibault Sottiaux: It's just like the intact, uh, and it is kind of like, you know, get the joy, um, you know, it's like preserve the joy of like programming.
[00:37:19] Andrew Zigler: I love that advice. It's like build your, build your own toolbox and think about those tools that you bring. I, I, I find a lot of the same luck. I, I, you know, I also use skills. I have a lot of customized things that do little weird, quirky things for me, or do it in a specific way. I prefer. Fur and like I love those.
[00:37:34] Andrew Zigler: I always pull those out with the first thing when I jump into something and I couldn't agree with that advice more. I love the Pokemon analogy. I'll push it even further and say it's more like, it's more like a chef in a kitchen like you, you have your knives, you bring your knives to work, you sharpen your knives.
[00:37:50] Andrew Zigler: You take care of your knives. Your knives, or. Tools and developers can think about their skills and how they work with their agents the same way. It's like that's your bag of knives. You could fold it up and take it with you. [00:38:00] You can make it sharper and better and bring it to the next thing. That's how everyone should be thinking about building the primitives that do plug into the system.
[00:38:07] Andrew Zigler: So that's really great advice. Tibo this has been a great look at, you know, behind the curtain at OpenAI. It's been a pleasure to have you on the show. Um, before we wrap up is, you know, is there anywhere you'd like to point our audience to, to go learn more about Codex? Go follow you and the work you're doing?
[00:38:23] Thibault Sottiaux: Yes, you can follow me on Twitter. Fairly active there, sharing tips, um, almost daily. lot of my team is also active. Then we have increasingly amazing developer documentation that is quickly coming together. And then of course, like the open source repo where, you know, you can follow issues, contribute to this discussion and, you know, be part of the community. And I really want to thank, uh, you know, thank you for having me on the show today. Uh, it's been like really a pleasure to talk about all these things.
[00:38:52] Andrew Zigler: Yeah, amazing. It's been great to have you too. And we're gonna include the links to all that stuff in our show notes. People can go check it out. And for our listeners, you know, that's it for this [00:39:00] week's Dev Interrupted, but if you wanna go deeper on how agentic AI is specifically reshaping engineering leadership.
[00:39:06] Andrew Zigler: Definitely check out our LinkedIn and Substack newsletters. Just search for Dev Interrupted anywhere. Uh, if, if you're listening to this, be sure to check out the included newsletter as well. We share analysis and takeaways from engineering leaders like Tibo each week and we're gonna continue to follow this story.
[00:39:21] Andrew Zigler: And Tibo, thanks again for joining us. It's been so fun to have OpenAI on the show, and we'll have you back some point soon I'm sure. Take care.
[00:39:29] Thibault Sottiaux: Of course. Bye.
