TestingCatalog says it found unreleased interface code inside OpenAI’s platform for a hosted Managed Agents product, surfacing just weeks before the company’s DevDay on September 29 at Fort Mason in San Francisco. The outlet is explicit that this is a codebase sighting, not an announcement. OpenAI has said nothing public about the feature, and functionality found this way routinely gets rewritten or dropped before it reaches users.
What TestingCatalog says it observed directly in the interface: screens for creating agents and environments, managing sessions, and enabling skills and plugins, plus an entry point through an OpenAI Developers Plugin and support for self-hosted environments. That is a materially different kind of claim from the ad-agent functionality discussed below, which the outlet frames as inference rather than something it saw rendered on screen.
What TestingCatalog describes looks a lot like Anthropic’s Claude Managed Agents, a hosted product whose pieces include subagents and audit logs, credentials the platform manages, permissions scoped per task, sandboxes, harnesses and sessions. Whether Anthropic’s version counts as a proven success is unclear even to TestingCatalog. Enterprises use it, but the outlet notes many developers still choose cheaper alternative stacks over the managed offering.
A managed agent product is a bet that companies would rather rent the whole loop than assemble it themselves, the same wager cloud providers made when they sold infrastructure instead of servers. For OpenAI, entering that market puts it in competition not just with Anthropic but with the harness and orchestration vendors that have built businesses on top of its own API. That dynamic barely existed when the company shipped custom GPTs in November 2023.
That earlier product history matters for judging how durable any DevDay announcement would prove. Custom GPTs were succeeded by an Agent Builder in October 2025, and that in turn is being retired after November 30, with the Agents SDK and Workspace Agents taking over, per an OpenAI update from June that TestingCatalog cites. Three product generations in under three years is a fast churn rate for anything a business is meant to build on top of.
TestingCatalog frames price as the open question that will determine whether Managed Agents gains traction where Anthropic’s version has not fully broken out. Strong computer-use capability in OpenAI’s newer models gives the company a technical case for undercutting on price or winning on capability. The outlet does not claim OpenAI has settled on pricing, and neither do we.
The advertising angle is the part of TestingCatalog’s report that carries the least evidence. The outlet reports work on a path where an advertisement inside ChatGPT hands the reader to an agent rather than a normal reply, so the agent walks them toward the purchase instead of simply sending them elsewhere. TestingCatalog itself calls the competitive threat to Meta and Google speculation, and that hedge should be read literally. This is analysis of what the capability could mean if built and shipped, not confirmation that OpenAI plans to deploy it that way. OpenAI already runs ChatGPT Ads and Ads Manager, and an August update said it would keep exploring native ways for businesses to reach consumers, which is the only on-record company statement TestingCatalog ties to the idea.
Operators evaluating agent infrastructure should treat this as a reason to watch DevDay on September 29, not a product to plan around yet. If OpenAI does introduce Managed Agents that week, the first fact worth checking is pricing relative to Anthropic’s offering, since that is the variable TestingCatalog identifies as the one OpenAI still has to prove it can win.
TestingCatalog reported the findings on September 7, 2026.
