OpenAI managed agents surface in code before DevDay

OpenAI's DevDay is September 29, 2026 at Fort Mason in San Francisco, and the keynote is free to watch. What ships on stage is not confirmed. But TestingCatalog reported today that OpenAI's shipped client code carries references to Managed Agents — creating agents, configuring environments, enabling skills and plugins, with a self-hosting option in the mix. Strings in a build are a reliable signal of intent and an unreliable signal of timing. The part worth acting on is not the leak. It is that three frontier vendors are now converging on the same abstraction, and you get to decide where your agent definitions live before one of them decides for you.

What actually happened

The confirmed facts are thin and easy to state. OpenAI DevDay runs September 29 at Fort Mason, in-person applications are closed, and the keynote is livestreamed. Everything about Managed Agents comes from TestingCatalog's read of the codebase — not an OpenAI announcement, not a press release, and not something we would price a decision on.

The confirmed context is more interesting than the rumor. Anthropic shipped the same shape three days ago: ant CLI v1.30.0 added ant apply on September 3, which creates and updates agents, environments, skills, memory stores, and deployments from files in your repo, and writes a claude-lock.json to track versions. Google already sells managed agents with token caps and lifecycle hooks. If OpenAI lands a third version of this at DevDay, the category stops being a differentiator and starts being table stakes.

Why this matters for your business

A managed agent runtime is genuinely useful. Someone else runs the container, handles retries, meters the spend, and gives you an audit log. For a small team that is real leverage, and we are not going to pretend otherwise.

The question is narrower than build-versus-buy. Where does the agent definition live?

There are two answers and they cost very different amounts. If your agent's system prompt, tool list, permission scopes, and environment config live in a vendor console — clicked in, stored server-side, exportable as nothing — then switching vendors means re-authoring every agent by hand, from memory, under deadline. If those same things live in your repository as files that a CLI applies, switching means rewriting an adapter. One is a migration project. The other is an afternoon.

Anthropic's ant apply is notable precisely because it picks the second answer: your agents are files, diffable in a pull request, reviewable by a human who is not the person who wrote them. We wrote about that when it shipped. Whatever OpenAI announces on the 29th, hold it to the same test.

Three things to do before DevDay, none of which depend on what gets announced:

• Put every agent definition in version control today, even if the vendor console is currently the source of truth. Export it, commit it, treat drift as a bug.
• Keep tool access behind MCP servers you run. The runtime can be rented. The connection to your database, your CRM, and your payment records should not be re-plumbed every time you change models.
• Write down the switching cost. Count the agents, the tools each one calls, and the hours to re-author. If that number is uncomfortable, it is the number a vendor's pricing page will eventually be built against.

Managed runtimes are a fine thing to rent. The agents themselves are yours, and they should be readable without logging in to anybody.

Sources: OpenAI DevDay 2026, TestingCatalog, Claude Platform release notes.