OpenAI Opens Agents API Beta With a Managed Codex Harness
The September 10 beta provides durable sessions, progress streaming and sandbox choices while OpenAI manages orchestration, context compaction and recovery.
Edited by Tyronne Panaino
OpenAI released the Agents API in public beta on September 10, 2026, according to its dated release record. The update gives developers a managed Codex harness for building agents, rather than requiring them to operate that harness themselves.
The API changelog says OpenAI handles session orchestration, context compaction and recovery. The announcement is therefore about the infrastructure around agent work, not simply the availability of another model.
Durable sessions preserve a continuing workflow
The official records describe durable sessions that can continue across turns, stream progress and connect developers' own tools and MCP servers. These are relevant to applications whose work needs to continue after an initial request, with progress available along the way.
The practical distinction is between selecting a model and deciding who operates the machinery that keeps its work moving. The beta offers a managed option for that machinery. It does not establish that a developer's connected tool is correct, that its permissions are appropriate, or that a completed task meets the application's requirements.
The harness does not dictate one sandbox location
Developers can run agents in OpenAI-hosted sandboxes, connect infrastructure of their own, or use a supported provider. That choice separates the managed harness from a single mandatory execution environment.
For a team evaluating the beta, the next useful questions are where code and files will be handled, what tools the application will expose, and how results will be checked. These are evaluation questions, not claims that every environment offers identical controls.
Public beta is the current release boundary
The retrieved records label the service a public beta, not general availability. A meaningful next checkpoint is a small controlled workflow that tests continuation, progress visibility and the chosen sandbox arrangement before expanding the application's authority. No production reliability or performance result is claimed here.
Status
Confirmed official beta release. Internal confidence is medium: the release record and changelog are both OpenAI sources and do not provide independent operational corroboration.
Sources
Update note: Reviewed against the clean official release records in this run. Beta stability and application-specific behavior remain untested.
Sources
- OpenAI — Release notes — official
- OpenAI API — Changelog — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.