AN EXECUTIVE VIEW ON AIWe speak with executives and industry leaders about how their companies are deploying artificial intelligence and putting it to work. In this edition, Tech in Asia's Melissa Goh speaks with Thibault Sottiau, head of Codex at OpenAI, about the rise of AI coding agents and what they could mean for software development. The interview has been edited for brevity and clarity.  Codex recently crossed 4 million weekly active users, up from 3 million just weeks ago. Where is that growth coming from? Are startups driving this, or are enterprises becoming the bigger story? It's both, but for different reasons. Startups and individual developers moved first because Codex gives them immediate leverage. You can go from an idea to shipping something much faster, and that matters a lot when you're operating alone or in a small team. But what's becoming interesting now is the enterprise side. The way I think about Codex is that we are building an agent first, then figuring out the best places to put that agent to work. I keep coming back to Asia Pacific as an example, because the pace of adoption here is incredible. Weekly active users in Asia Pacific grew 11x, faster than our global growth. And if you look at Singapore alone, it's already a top five market globally for Codex adoption and engagement. There are plenty of examples emerging in the region, but one that's top of mind for me is Sea Limited. They've rolled out across their engineering organization, with OpenAI internal data showing 87% of Sea users are weekly active users. So I would not frame it as startups versus enterprises. Startups showed how quickly Codex can change the speed of building. Enterprises are showing how it becomes part of the operating system for software teams, and eventually a pattern for how agents can support many kinds of work across a company. Have you seen any interesting, clever, or unique use cases emerge from users in Asia that surprised you? One example I really like is Hiroki Tomiyasu, a farmer in Japan. He does not come from a traditional engineering background, but he used ChatGPT and Codex to build practical tools for his own farm, from greenhouse controls to remote temperature management. What stands out to me is that Hiroki was already closest to the problem. Codex helped him move directly from understanding the frustration to building a working solution, without needing an external vendor or a software team. Codex is not just making development faster. It is making development accessible to more people, so real-world problems can be solved by the people who understand them best. Southeast Asia is especially interesting because the problems are very practical. Builders here often have to solve many things at once: different markets, multiple languages, and complex flows. That creates the kind of high-context environment where agents can be useful. There's a growing argument against using agentic coding tools. Some argue they create code bloat, more technical debt, and don't necessarily improve product velocity. What do you think critics are getting wrong or right? If the goal is simply to generate more code, then I think the criticism is justified. However, that's not how I think about Codex. To me, Codex is an agent first. The hard problems in software engineering are often about finding the right context, understanding the system, deciding what action to take, and validating the result. The product philosophy matters too. We do not want to keep adding layers of scaffolding that become fragile as models improve. The goal is to find simple primitives that let the agent use tools, follow instructions, and get better as the underlying model gets better. So the real question is not whether agents can produce code - of course they can. The question is whether teams use them to become better engineers with stronger workflows, better review loops, and a higher quality bar. What's the most unexpected thing you've personally seen someone build or do with Codex? Codex may have started as a coding story, but I love seeing how more knowledge workers are using the agent. It still surprises me how quickly people are treating it as a teammate they can delegate real work to. The unexpected part is not always the final app. It is the way the workflow changes. I often hear of business leaders looking at using Codex to build a "second brain." The boundary keeps moving and that's exciting. A designer, researcher, founder, or small business can now go from "I have an idea" to "I have a working prototype" much faster. Codex started in software engineering, but the behavior it teaches is broader - describe the outcome, give the agent context, review its work, and keep iterating. For a lean startup with maybe five engineers and no dedicated AI team, or no engineers at all, what's the "minimum viable workflow" today to get real leverage out of Codex instead of just generating more code? The minimum viable workflow is to stop asking Codex for isolated snippets and start giving it real work with context and review. Pick one concrete problem, explain the goal, share the relevant files or notes, ask Codex to propose a plan, and then let it make a small change that you can inspect. For a five-person engineering team, that might mean delegating bug fixes, test coverage, documentation, small refactors, or prototype branches. The human still owns judgment. Is this the right product decision, does the code match our standards, did the tests pass, and is this safe to ship? But Codex can do a lot of the context gathering, implementation, testing, and cleanup in the background. For a non-technical founder, I would start even more simply. Build a lightweight internal tool, a landing page, a tracker, a dashboard, or an automated workflow that saves you time every week. The point is not to become a software engineer overnight. The point is to learn the muscle of delegating work clearly by defining the outcome, giving examples, reviewing the result, and iterating. As someone leading Codex, has using coding agents changed the way you personally work or think about software engineering? Definitely. The important skill is not if I can type every line myself, but if I can define the right problem, give the right context, set the right constraints, and evaluate the result. Inside OpenAI, usage is everywhere. Every engineer uses Codex by default, and we're completing 2x more pull requests per engineer. We also ship much faster. Because we use Codex to build Codex, this has been the fastest improving product in OpenAI. Almost every single week we are launching new features. If you're having a hard time keeping up with our products, blame it on Codex! That changes how our teams work too. Specifications, tests, repo conventions, review comments, and product judgment become even more important because they are the way you steer the agent. The better you are at describing what good looks like, the more useful Codex becomes. It has also made parallel work feel more natural. Instead of one person moving one task forward at a time, you can have several agents exploring different implementation paths, writing tests, investigating issues, or preparing a review. The engineer's role becomes higher leverage but also more responsible. What's an underrated feature or behavior inside Codex that power users understand, but most people completely miss? Power users understand that context and instructions are the product surface. Codex is much more useful when you teach it how your team works. For instance, your repo conventions, how you like tests written, what patterns to avoid, what "done" means, and which workflows should be repeated consistently. That is why things like Skills, automations, memory, and repo instructions matter. They let teams turn a good one-off interaction into a repeatable workflow, so Codex behaves consistently across people and projects. The other underrated behavior is asking Codex to review and validate, not just create. Some of the highest-value use cases are debugging, test generation, code review, finding edge cases, and explaining a complex system. If you only use it to produce new code, you miss a large part of what makes agents useful. On a personal level, what's something about the future of AI coding tools that genuinely excites you, and something that quietly worries you? What excites me is the access story. Codex is available to everyone who uses ChatGPT, which means more people can start building with agents where they already are. Software has always been one of the most powerful ways to turn an idea into something real, but for many people the barrier was too high. Agents lower that barrier. That is especially meaningful in regions like Southeast Asia, where builders often work across different markets, languages, and constraints. If agents help more people move from using AI to building with AI, the ecosystem gets much more interesting. What worries me is people confusing more output with better engineering. If teams use agents only to generate more code without stronger review, testing, security, and product judgment, they can create more complexity rather than more value. The future I want is not one where software teams ship blindly faster, but one where agents help people build with more discipline, better context, and higher quality. What does a typical day look like for you now? How much of it is already shaped, managed, or automated by AI agents? It is less "AI runs my day" and more "agents increasingly handle the surrounding work so I can spend more time on judgment." A typical day still involves product decisions, team discussions, roadmap tradeoffs, customer and developer feedback, and a lot of decision-making. But more of the work around those decisions is agent-shaped now. Codex can help investigate an issue before a discussion or summarize context across a codebase. So instead of switching constantly between finding context, writing code, reviewing diffs, and documenting decisions, I can delegate more of those pieces and stay focused on the direction and quality of the work. Agents do not remove the need for taste, judgment, or responsibility - they raise the leverage of those things. I think the people and teams who get the most from Codex are the ones who learn how to direct agents well and still keep a high bar for what ships. |
