Romain Huet: agents primary developers

Software development is starting to split in two. Humans still decide what to build, but more of the actual clicking, reading, writing, and testing is getting pushed onto agents like Codex. That is the backdrop for Romain Huet’s point this week — the head of developer experience at OpenAI said agents are becoming the “primary developers” in API ecosystems, and that they are “even hungrier for good DX than developers themselves.” The comment lands because OpenAI has spent the last year building exactly the stack that makes this possible — Responses API, Agents SDK, and Codex as a coding agent that can run tasks across tools and files. (vivatech.com) ### What is he actually claiming? He is not saying human engineers are gone. He is saying the first reader of your docs, the first caller of your API, and the first system trying to stitch your product into a workflow may increasingly be an agent. That changes the center of gravity. “Developer experience” used to mean helping a person get from docs to hello world. Now it also means mak(vivatech.com)rrors, and keep going. (developers.openai.com) ### Why does DX get more important, not less? Because agents are powerful, but brittle in boring ways. They need clean schemas, predictable auth flows, stable error messages, explicit rate limits, and docs that say exactly what a tool does. A human can guess around a messy edge case. An agent often cannot — or it can, but only after wasting tokens, time, and money. That is the point inside (developers.openai.com)g polish and start acting more like control surfaces. (developers.openai.com) ### Why is Codex the obvious example? Because Codex is no longer just a code-completion toy. OpenAI describes it as a coding agent that runs locally in the terminal, and also inside IDE workflows, to take on longer-horizon tasks. OpenAI’s own recent materials frame Codex as moving from “code generator” toward a “teammate” that helps with planning, architecture, debugging, and documentation. (developers.openai.com)aving less like autocomplete and more like an operator moving through an API ecosystem. (github.com) ### What changed over the last year? OpenAI has been steadily shipping the plumbing for agentic development. In March 2025 it introduced the Responses API as a core primitive for agentic applications, with built-in tools like web search, file search, and computer use. The Agents SDK gives developers a higher-level runtime for orchestration, state, approvals, tracing, and multi-step workflows. More recently, OpenAI added s(github.com) long-running agents. Basically, the company has been turning “agent” from a demo word into a platform shape. (openai.com) ### So what should API teams hear in this? They should hear that their customer is now partly human and partly machine. The old checklist — decent docs, quickstarts, SDKs, examples — still matters. But now the machine-readable layer matters just as much: structured outputs, explicit tool definitions, testable examples, and fewer ambiguous steps. If an agent is the one integrating your product, ever(openai.com)the real strategic shift. (developers.openai.com) ### Does this change the role of the human developer? Yes, but mostly upward. Humans spend less time typing boilerplate and more time picking constraints, reviewing outputs, setting guardrails, and deciding whether the system is solving the right problem. Huet has been making this broader case in talks around “vibe engineering” and agent-assisted software work — the job moves from line-by-l(developers.openai.com)y works when the tools underneath are reliable enough for delegation. (forum.openai.com) ### What’s the bottom line? Huet’s comment sounds like a slogan, but it is really a product thesis. If agents are becoming the main actors in software integration, then developer experience is no longer just about making developers happy. It is infrastructure for machine action — and the companies that treat it that way will be easier for both humans and agents to build on.