Published: 2026-09-10
Sketch-to-layout in ChatGPT Image 2.5, and a real test of edit consistency
Codex Desktop for Business: Setup, Approval Modes and Scheduled Tasks
Docker's Agent Sandbox Runtime: Scoped Credentials Per Task, Local to Cloud
Codex browser and computer use: saving flows as skills you can schedule
Same prompt, three harnesses: a Claude vs Grok vs Codex writing test (hands-on comparison, not a how-to)
Progressive context shaping: steering agent runs that last ten hours
OpenAI's Agents API is a hosted Codex harness — here is what it takes over
Chapters / key moments (click to jump — plays here on the page)
The pitch is a list of the things you currently maintain and would rather not. Getting a long-running agent into production means connecting tools, tracking progress, managing context, and securing and maintaining infrastructure around it — and the Agents API brings a hosted version of the Codex harness to your application, with OpenAI handling orchestration, sessions and context management. Given that the harness is roughly 80% of an agent, a vendor offering to run it changes the build-versus-buy calculation for anyone currently maintaining their own loop.
Source video
"Introducing the Agents API" by OpenAI — Watch on YouTube →
Gotchas & Caveats
- This is OpenAI's own launch video — it is an explainer of intended capability, not an independent evaluation, and nothing here has been tested by us or anyone else on camera.
- The incident-investigation example is a demonstration scenario. It is well chosen because it needs tools, long context and delegation all at once, but it is a showcase.
- No pricing, limits or availability detail appears in the video; see our 2026-09-11 changelog entry for the release context.
- Public beta as of 2026-09-10. "Hosted harness" also means a dependency on someone else's harness — the trade-off against running your own is real in both directions, and this video only argues one side.
Key Takeaways
- It is a hosted harness, stated plainly. Not a model, not a framework — the Codex harness as a service, with orchestration, session handling and context management on OpenAI's side.
- You still control the execution environment and the tools. The sandbox can be OpenAI's, a third-party provider's, or your own infrastructure — which is the detail that decides whether this is adoptable in a regulated environment.
- Tools come in through MCP, and procedures come in as skills. The worked example connects observability data and recent code changes as MCP servers, and supplies the team's outage investigation runbook as a skill — the same instructions-as-a-first-class-input pattern that LinkedIn arrived at independently.
- Programmatic tool calling is the token argument. Faced with more logs than fit in a context window, the agent processes and filters them in code — so tokens are spent on the information it needs rather than on passing raw data around. This is the concrete answer to "how does a hosted agent handle data at production scale".
- Multi-agent orchestration for independent work. In the example, one subagent examines recent changes, another checks telemetry, and a lead agent unifies the findings.
- Compaction keeps a single long session viable by summarising prior work so the agent can continue — the alternative to fanning out when the work is not actually parallel.
- The output is a shareable artifact, not a chat log: likely root cause, supporting evidence and suggested next steps bundled into a single file for the on-call team to review.
More ChatGPT news
Sketch-to-layout in ChatGPT Image 2.5, and a real test of edit consistency
Codex Desktop for Business: Setup, Approval Modes and Scheduled Tasks
Docker's Agent Sandbox Runtime: Scoped Credentials Per Task, Local to Cloud
Codex browser and computer use: saving flows as skills you can schedule
Same prompt, three harnesses: a Claude vs Grok vs Codex writing test (hands-on comparison, not a how-to)
Progressive context shaping: steering agent runs that last ten hours