OpenAI introduced the Agents API on 10 September 2026. The company says it is the same harness that powers OpenAI Codex, now reachable through a public-beta API. OpenAI operates the managed harness. The core logic is also visible in the open-source Codex codebase.
The buying decision is not “which model string.” The launch example uses GPT-6 Astra. The decision is whether you want OpenAI to keep the session, compact context, steer mid-run, and spawn subagents, or whether you keep that layer on Cursor, Devin, GitHub Copilot, or a stack you built.
The API is four objects, not a chat wrapper
The Agents API overview defines the product around four objects. Treat these as the contract, not as marketing synonyms for a thread.
- Agent: the model, instructions, tools, and MCP servers available to the run.
- Environment: an optional sandbox or computer where the agent reads files, loads skills, and runs commands.
- Session: a durable instance that works on tasks and can continue after a turn ends.
- Events and items: the inputs you send and the output the session produces.
OpenAI says the managed harness runs commands in a sandbox, applies skills, connects tools or MCP, lets you steer while it works, summarizes earlier work to stay inside the context window, delegates to subagents, and resumes a session where it left off. That is a platform job. It is not an inline autocomplete job.
Hosted, self-hosted, or a partner box
You pick the compute. OpenAI hosts a sandbox that it says uses the same infrastructure as Codex and ChatGPT. You can also run a self-hosted environment, or use a partner listed on the launch post: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.
- Hosted: OpenAI provisions the box. Fastest path. You inherit OpenAI’s container rates and its data controls.
- Self-hosted: you keep files and commands on your infrastructure. You still talk to the Agents API.
- Partner: first-class integrations for VPC deploy, storage, and different CPU or GPU profiles.
US residency only. Self-hosted does not buy you ZDR.
The overview is explicit. The Agents API currently supports data residency only in the United States. It does not support Zero Data Retention. Choosing a self-hosted sandbox does not make the API ZDR-eligible. OpenAI points to its platform data-controls pages for the retention detail.
How this sits next to Cursor, Devin, and Copilot
The September coding-agents shortlist still stands. Agents API does not replace an IDE. It replaces (or competes with) the session layer those products also sell.
- Cursor now has Projects: a coordinator that keeps cloud agents on a body of work after you close the laptop.
- Devin and Devin Desktop sell a worker plus a desktop. See Windsurf is Devin Desktop if your tabs still say Windsurf.
- GitHub Copilot is adding admin gates on agent operations. That is a permission product, not a public session API.
- A self-built harness keeps you on raw model calls. You own compaction, resume, and subagents. You also own the on-call.
OpenAI’s launch page quotes customers. Ciridae’s CTO says an evaluation score moved from 0.71 to 0.85 and cites a 4x latency cut. SafetyKit’s staff says cost per case fell 60 percent after a migration. Those are quotes on OpenAI’s page. They are not an independent audit. Run your own tasks.
What to ask OpenAI this week
- Confirm in writing that residency is US-only and that ZDR is unavailable even with a self-hosted sandbox.
- Get the current token, tool, and hosted-container rates for the model you will actually run (Astra in the launch example).
- Name the sandbox path: hosted, self-hosted, or a listed partner. Ask who can see files and logs.
- Ask whether MCP servers, vaults, and skills you already run on Codex transfer without a rewrite.
- Ask how session deletion and artifact deletion work, and who can pull a transcript for an incident.
- Put the same questions to Cursor Projects and Devin so the answers sit on one sheet.
Astra is a new model string. It is not a freeze on retirements. Keep the 2026 deprecation-wave playbook next to this page, and start tool-using runs from the agent-security checklist.