STRONG SIGNAL: OpenAI looks set to host the whole agent, not just the model

Every time OpenAI shifts its agent strategy, it leaves a trace in its own code before announcing anything. That is what happened with Agent Builder in 2025, and it is what TestingCatalog says it found again: unfinished screens inside OpenAI's platform for a product called Managed Agents, which creates and manages agents, execution environments, sessions, and a skills and plugins system wired through an OpenAI Developers Plugin. The find surfaced thirteen days before DevDay, set for September 29 in San Francisco, and spread across at least three independent publications describing the same interface elements.

Why this differs from running an agent on your own infrastructure

Today, building an agent with OpenAI means orchestrating the execution environment yourself: containers, credentials, observability, all on your own using the Agents SDK. What the code suggests is a layer shift: OpenAI hosting both the agent and the environment it runs in, including an option for self hosted environments for teams that do not want to hand everything over. That is exactly the proposition Anthropic already sells as a separate product, and the comparison is not a rhetorical flourish: the same reports describing the find say the interface closely mirrors what Anthropic offers.

The trail behind the trail

Managed Agents does not come out of nowhere. OpenAI launched Frontier in February to govern agents at scale, made Workspace Agents generally available in April, and shipped Presence in July aimed at agents already running in production inside client companies. Placed on that timeline, a product that hosts both the agent and its environment is the obvious next step, not a detour. That does not confirm the code find, but it explains why it is plausible and why it surfaces right now.

The Agent Builder clock is what gives this story urgency

There is a practical reason this matters beyond curiosity: Agent Builder, launched in October 2025, was marked for deprecation in June and shuts down on November 30, with Evals going read only on October 31. Anyone who built on Agent Builder is on a clock regardless of any DevDay announcement. If Managed Agents is real and gets presented as the natural successor, the migration path becomes far clearer. If it does not show up, the deprecation pressure stays exactly the same, just without an obvious destination.

What is still unconfirmed, and what changes if it lands together

None of this comes with pricing, HIPAA availability, usage limits, or a confirmed launch date. Running in parallel, with weaker evidence, is the idea of an in ChatGPT ad that hands off directly to an agent that completes the purchase, and the same reports treat that part as a read of what the capability could mean, not something seen rendered on screen. MaxAssistant's read is direct: the interface find is solid enough to call a strong signal, but the ad-to-agent product stays in hypothesis territory until there is a screenshot, code, or OpenAI confirmation. Worth watching DevDay on September 29 to see which of the two pieces the company decides to show.

Sources

TestingCatalog via AI Insiders, Managed Agents code find: https://aiinsiders.net/article/leaked-code-points-to-openai-managed-agents-ahead-of-devday Claypier, interface breakdown and OpenAI agent product timeline: https://claypier.com/en/openai-devday-2026-managed-agents/ ByteIota, Agent Builder and Evals deprecation schedule: https://byteiota.com/openai-devday-2026-managed-agents/ CryptoBriefing, Frontier, Workspace Agents and Presence timeline: https://cryptobriefing.com/openai-managed-agents-devday-2026/