On September 10, OpenAI put the harness behind Codex into a public API. The Agents API, now in public beta, hosts the session, the context compaction, and the coordination of subagents. You choose where the agent runs: an OpenAI sandbox, your own infrastructure, or a partner environment. There is no extra fee for the harness. You pay for the tokens and tools it uses.

That is a real product shift. It is not a workflow. The announcement says OpenAI maintains the layer that keeps a long-running agent from losing the thread. It does not say the agent now owns your CRM, your inbox, or the person who gets pinged when a step dies at 11pm. Source: Introducing the Agents API, OpenAI, September 10, 2026.

What they took off your plate

If you have tried to keep an agent alive across a long task, you already know the boring half. Context fills up. Tool lists get huge. A second agent needs its own memory. Someone has to notice when the run stalls.

OpenAI is selling that half as infrastructure. Their docs say the API manages sessions, orchestration, compaction, and recovery. Your application supplies the tools and picks the environment. Partner sandboxes named in the launch include Cloudflare, DigitalOcean, Modal, Oracle, and Vercel, among others.

For a software team, that is useful. You stop rebuilding session plumbing every time the model changes. For an operator, the useful question is narrower: which part of last Tuesday's job did this replace?

The constraint that actually matters

The product page is the friendly version. The API docs are the version you should read before you move a customer file into it.

OpenAI's Agents API overview says data residency is currently United States only, and Zero Data Retention is not supported. Choosing a self-hosted sandbox does not make the API eligible for zero retention. Session state still sits with OpenAI. Source: Agents API overview.

That is not a footnote if the work touches client records, invoices, or anything a contract says must stay in a region or leave no trace. A self-hosted box for the code does not move the session. If that constraint blocks you, the honest move is to keep that path off this API. Do not treat the beta as a workaround.

A harness is not the path between your tools

The same week, a lot of coverage will treat this as "agents are now easy." Easy to start a session is not the same as easy to finish a job your team already does by hand.

A hosted harness can draft, search, and hand work to a subagent. It still needs someone to define:

  • What event starts the work, so it does not wait on a person to paste a prompt
  • Which system the result is written into, so the chat is not the database
  • What the agent is allowed to touch, and what requires a person
  • Who gets the failure, with enough context to fix it once

That list is the workflow. OpenAI just made the session layer a product you can rent. The list did not move into their API.

What to do with the launch

Do not rewrite the company around a beta endpoint. Do not ignore it either. The direction is clear: the scaffolding that used to be a custom engineering project is becoming a line item, billed as tokens and tools, with residency limits written in the docs.

If you already have one painful path, ask whether a hosted session helps the drafting or research step inside it. Then keep the trigger, the system of record, and the exception owner where you can see them. If the path handles client data and the docs say US-only with no zero retention, that path stays off the API until the controls match the requirement.

They will host the harness. You still own the path. Those are different purchases.