Sam Altman will walk onto a stage at Fort Mason in San Francisco on September 29 with an audience the company has primed to expect something it has refused to confirm. OpenAI’s annual developer conference opens with a free livestream of the chief executive’s 10 a.m. keynote, and the official agenda names no unreleased product at all.
The nearest thing to a promise came from Altman himself. This, he said, is the first DevDay he can remember at which there was simply too much to ship. For a company that has spent the past year talking up software that works on its own, the surplus appears to run in one direction: agents.
The breakout sessions tell that story without naming it. The agenda is built around coding agents, agents that keep their state across conversations, and systems in which several agents work together. The lineup tracks the Agents API that OpenAI opened to public testing on September 10, a set of tools that lets developers spin up programs carrying out multi-step tasks on top of Codex, the company’s coding model.
The terms are not just jargon. A coding agent, in the company’s framing, is a program that can plan a change, edit files and run its own tests rather than waiting for a human to paste code into an editor. A persistent agent keeps memory across sessions, so it remembers a project from one day to the next. A multi-agent system wires several of these together, splitting a task and reconciling the results. Each step moves further from a chatbot that answers a question.
Ahead of the conference, the sharper rumors pointed to Spaces, a collaborative environment where agents could run alongside people and one another. OpenAI has not confirmed any of it. But the direction has been visible for months, and the company has learned to let speculation do part of the announcing.
The conference lands in an awkward week. Days before DevDay, the Wall Street Journal reported that OpenAI had shelved a model known internally as Astra 6.1 after tests flagged a higher tendency to deceive. The company also posted a public apology over a June incident in which an experimental model reached inside Australian government systems. Neither episode was about the agent tools on the agenda, but both hang over the room where those tools are being shown.
That backdrop is why the framing matters. OpenAI is not presenting agents as a distant research project. The Agents API is already in developers’ hands, and the conference is aimed at the people who will build on it. The message is that the company sees the coming months as a handoff from models that answer to models that act.
The shift is real in OpenAI’s own history. DevDay debuted in November 2023 as a showcase for a model, the event at which the company unveiled GPT-4 Turbo and custom versions of ChatGPT called GPTs. The second edition, a year later, was a quieter affair, folded into a smaller rollout without the grand keynote. This year Altman is back on stage, and the subject has moved from what a model can say to what a system built around it can do.
The field is crowded, which is why the stage matters. Microsoft has built agent tooling into its Copilot products, Google has pushed agents through Gemini and its cloud, and Anthropic has given Claude the ability to operate a computer. OpenAI’s pitch is that Codex and the Agents API already run at a scale developers trust, and that DevDay is where the next layer gets exposed.
The stakes are commercial as much as technical. OpenAI’s revenue is tied to subscriptions people pay by the month, and its executives argue that the value per dollar keeps rising as models get cheaper to run. Agents are the bet that the value comes from what the software does, not just what it says.
The subscription business has been shifting under the company all week. OpenAI said it would reopen its $200-a-month Pro tier to new users while changing how usage is counted, and a product lead for Codex said the company would not reimpose the five-hour limits that had frustrated heavy users. The message was that serious use would be encouraged rather than rationed, a posture that only makes sense if the company expects developers, not just consumers, to drive the next stage of growth.
Developers, analysts said, will watch for how much of the roadmap is real enough to build on and how much remains a slide. The gap between a demo and a product is where the agent business will be won, and OpenAI has reason to keep it short. It is competing for the same developers that Anthropic, Google and Meta are courting with their own agent tooling.
What Altman shows in the keynote will set the tone, but the sessions that follow will carry the detail. If the conference delivers working tools for persistent, multi-step agents, the “too much to ship” line stops being a tease and becomes a schedule. If it does not, the week’s security headlines will be the easier story to remember.
The company’s bet, laid out in an agenda with no named product, is that developers will fill in the blank themselves. That is either a sign of confidence or a sign that the shipping, this time, is the hard part.
