OpenAI opened DevDay 2026 at Fort Mason in San Francisco on Tuesday with a keynote focused on one theme: getting humans and AI agents to work in the same room. The headline reveal was Spaces, a collaborative environment where users and agents jointly generate and refine content, shown live for the first time after weeks of rumors. Sam Altman delivered the opening keynote at 10 a.m. Pacific, streamed free to the public, with breakout sessions and demos filling the rest of the day at the company’s biggest annual developer event.
Spaces is the next step in a line OpenAI has been building for two years, and the lineage matters for understanding what it actually is. Canvas arrived in 2024 as a shared editing surface for documents, the first time users and the model manipulated the same text side by side. Shared projects followed in 2025, widening the idea from one file to whole workflows with multiple documents and shared context. Then came Workspace Agents, powered by Codex, which reached general availability across ChatGPT Business, Enterprise and Edu plans in mid-2026 and plugged into platforms like Slack to run inside existing team tools. Spaces folds those threads together: instead of an agent being summoned to help when called, it sits in the workspace alongside the people, editing and commenting like another collaborator with its own account of what the work is.
Agents are the product now
The session agenda made the direction explicit. Breakouts covered coding agents, persistent agents that keep running between sessions, and multi-agent systems where several models divide a task among themselves. This follows the September 10 public beta of OpenAI’s Agents API, which runs agents on the Codex harness and supports long-running sessions, hosted or external sandbox environments, custom tools via MCP, and parallel subagents. In practice that API moved orchestration from developer-managed code into OpenAI’s managed layer, and Spaces extends the same logic to the human side of the workflow. The question OpenAI is putting to developers is how much operational work they are willing to hand to a managed runtime, and everything on stage Tuesday suggested the company wants the answer to be most of it.
Reports ahead of the event suggested OpenAI had prepared an unusually large batch of launches, with some outlets counting more than twenty possible product announcements across models, APIs, subscription plans and hardware. Altman himself fueled that expectation in the run-up, writing that getting ready for this DevDay was the first time in OpenAI’s history he had said there was too much stuff to launch. Not everything rumored materialized on stage, and the company has not confirmed a consumer product some outlets had previewed under a different name. What was confirmed is a continued push to make agents a normal part of how work gets done rather than a demo category that impresses at conferences and stalls in production.
The competitive frame
DevDay lands at a crowded moment. Anthropic and Google are both courting the same developers with their own agent tooling, and the race has moved from model quality to infrastructure: who hosts the runtime, who handles the sandboxing, who bills for the subagents. OpenAI’s bet is that owning the collaboration surface, the place where a person reviews and steers an agent’s work, is as valuable as owning the model underneath it. Control of that surface also means control of defaults, and defaults in developer platforms tend to decide markets more often than benchmarks do.
The safety debate around agents has also grown loud enough to shape the pitch. Nvidia shipped an entire Open Agent Safety Platform with Microsoft, Anthropic and roughly a hundred partners last week, citing reports from OpenAI, Meta and Google of agents escaping their sandboxes in testing. Against that backdrop, OpenAI pitching collaboration rather than autonomy reads as a deliberate choice. An agent that works visibly alongside a person, showing its edits in a shared space, is an easier sell to enterprise buyers than one left alone overnight, and it gives OpenAI a talking point that does not depend on claims of raw capability. The company has its own history here: Altman has previously framed gradual deployment, shipping agent features to limited audiences and expanding as behavior is observed, as the safest path available.
Pricing and access details for Spaces were not immediately laid out in the keynote materials reviewed before publication, and OpenAI did not respond to a request for comment ahead of this article. The company said recorded sessions will be posted on its site shortly after the event, so developers not in San Francisco will see the full picture within days. DevDay Exchanges, the regional spinoffs announced earlier this year, are scheduled for Bengaluru, Tokyo, Seoul, Paris, Berlin, London, Sao Paulo and Mexico City, extending the developer tour well beyond the keynote hall.
For the developer audience, the practical takeaway is that OpenAI wants to own the layer where agents and people meet, the way it earlier owned the layer where apps meet chat. Whether Spaces becomes that surface depends on details still to come: how it prices agent time, how it handles enterprise permissions, and whether the collaboration model holds up outside demos. But the direction of the company’s developer story is no longer in doubt. Everything on stage Tuesday assumed agents are here to stay, and the product question is only where they sit while they work.