OpenAI opens Codex's long-running harness to developers through the Agents API
OpenAI has put the Agents API into public beta, giving developers access to the managed harness behind Codex and ChatGPT. The service keeps long-running sessions alive, compacts context, recovers interrupted work and can let an agent execute code, read files, call tools or divide a job among other agents. Workloads may run in OpenAI-hosted sandboxes or customer infrastructure, although the beta currently limits data residency to the United States.
Artificial Intelligence··Night
The Codex harness enters public beta
OpenAI has moved the Agents API into public beta, opening the managed agent harness used inside Codex and ChatGPT to application developers. Rather than forcing a long job into one request-and-response cycle, the interface runs it as a persistent session. OpenAI handles those sessions, coordination between agents, compression of a growing context and recovery after an interruption. THE DECODER independently reported the release as a public beta that gives developers access to infrastructure operating behind Codex and ChatGPT.[1], [2]
Sessions can use tools and divide work
An Agents API session can remain active for hours. An agent may execute code in a sandbox, read files, call developer-defined tools or connect to Model Context Protocol servers. Model Context Protocol is an open communication scheme that gives AI applications a common interface to external tools and data sources. A lead agent can divide a large assignment into smaller jobs, hand them to other agents and collect the results inside the same session. Context compaction is used to keep earlier steps available without letting a long run exceed the session's working limit.[1]
Execution is flexible, but data limits remain
Agents can run in sandboxes supplied by OpenAI or on infrastructure provided by the customer. The public beta also names Cloudflare, Vercel and Oracle as additional execution environments. There is no separate charge for the interface itself; model, tool and hosted-environment use is calculated under their normal rates. Data controls are narrower in the beta. The documentation says data residency is supported only in the United States and that zero data retention is unavailable. Those limits directly affect an initial deployment decision for organisations subject to regional or regulated-data requirements.[1]
