Scaffolding is coping not scaling, and other lessons from Codex | OpenAI’s Thibault Sottiaux

**SPEAKER_1** (0:05)
My guest today is Thibault Sottiaux from OpenAI's Codex team. Thibault has been working shoulder to shoulder with research and engineering in San Francisco to solve one of the hardest problems in tech, true agent tech autonomy. Thibault, it's really great to have you here on Dev Interrupted.

**Thibault Sottiaux** (0:21)
Great, awesome to be here. Really excited to talk about all of this.

**SPEAKER_1** (0:25)
Great, so today we're going to look at the bitter lesson of AI development. 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 going to 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. But I want to start at Codex, just in your own words, like learning about OpenAI's flagship coding agent. Can you tell us a bit about what it is and why you describe it as an agent first instead of a product?

**Thibault Sottiaux** (1:02)
Yeah, so we think about the two parts. First and foremost, we're building a SOTA agent that is able to act and perform incredible amounts of work on 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. 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. 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, is like you find a remarkable amount of places where this agent comes in handy and can actually do economically valuable work. We're also thinking about what does this mean beyond just coding? Even for software engineers, it's not just about code generation, it's about solving many other parts of the day-to-day that are actually the bottlenecks.
Building that general agent is what we're after.

**SPEAKER_1** (2:01)
Yeah, and we're going to talk more about what those other bottlenecks are. We talk about that a lot on Dev Interrupted, especially in the last year, about how agentic development has really made bear all of those human and communication problems that are affecting teams really deeply. Something that you said there about 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. As an agent, it seems to be emerging as something that operates the same way. 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.

**Thibault Sottiaux** (2:41)
Yeah, that's right. If you think about building an autonomous, ever more capable entity and keep your mind flexible about what the best form factor is, what the best product around it is, and perhaps we put it to work inside our own products, but we also partner with other companies to put it to work in their products. It opens the door for a lot of great ideas that you don't necessarily have. We don't have them ahead of time. We can focus on building the agent and then figuring out where to put it to work later.

**SPEAKER_1** (3:13)
And what is it like to work in an environment where you're building a coding agent, 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?

**Thibault Sottiaux** (3:25)
That's right, 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 research influences the entire engineering roadmap for how we build out the agent. And one of the things that you can do when you're vertically integrated is you can decide where you actually fix problems. So you don't have to fix everything in your harness. Some of the things we decide to fix downstream, by training new models. And we know that by training the model, we will have a jump in the capability that we need, like three months down the line, six months down the line. And it allows us to do these trade-offs that you can't do without a vertical integration. There's also this thing called the No Free Lunch theorem, which is basically if you're trying to adapt and be intelligent in any possible distribution, well, this is going to be strictly less optimal that if you were to actually build something for a very specific distribution. And so by coupling the harness and the model, and that's what we named the agent, we're able to get a lift in capabilities. And that's obviously very interesting for something as important as coding.