OpenAI Opens the Codex Harness to All With the Agents API

How much of an AI agent is actually the model? Less than most teams expect. The larger share is plumbing: keeping a session alive for hours, trimming the conversation before it overflows, retrying a tool call that failed, resuming after a crash. OpenAI has now put that plumbing behind a single API. The Agents API, launched in public beta on September 10, exposes the same harness that runs Codex and ChatGPT. There is no platform fee; you pay for tokens, tools and container time at the usual rates.

That sounds like a convenience. It is also a bet on where the value in agent development sits, and the answer OpenAI is giving is: not in the harness you wrote yourself.

Four objects, one call

The design rests on four pieces. An Agent bundles the model, instructions, tools and any MCP servers. An Environment is an optional sandbox where the agent can read files, load skills and run commands. A Session is a durable instance of the agent working on a task. Events and Items are what you send in and what streams back. You describe all four in one request, and OpenAI runs the loop.

What the harness does on its own reads like the backlog of a typical agent team:

• Context compaction when a session nears its token limit
• Tool search, so tool definitions load only when needed instead of on every request
• Programmatic tool calling, with parallel execution and chained results
• Subagents with their own context, capped by a max-concurrent setting

Tools cover MCP servers, custom functions, built-in web search, vaults for secrets and webhooks for events, with tracing and usage monitoring built in. The sandbox can be hosted by OpenAI or by you. On the self-hosted side the partner list is long: Cloudflare, Vercel, DigitalOcean, Oracle, Modal, E2B, Daytona, Runloop and Blaxel. The documentation examples run on GPT-6 Astra.

So is the Agents SDK obsolete?

Not according to OpenAI's own comparison. The Responses API leaves everything in your environment, including history management, and carries the highest integration effort. The Agents SDK runs inside your application with storage you manage, a medium lift. The Agents API hands the harness and the saved session state to OpenAI, the lowest lift. The trade is control for speed, and each tier still has a use.

The two lines worth reading twice

The beta notes carry two restrictions that matter more than any feature. Data residency is US-only for now. Zero Data Retention is not supported, and running the sandbox on your own infrastructure does not change that. Session content lives on OpenAI's side for a while whether you like it or not.

What a business should actually do with this

For prototypes, internal tooling and code-heavy work, the Agents API removes weeks of infrastructure work. A document-processing agent or a legacy-code migration bot can now be stood up in days, and the saved session state means a long task survives a dropped connection. That is real value, and the pricing makes it easy to trial.

For anything touching customer records, contracts or employee data, the two restrictions above decide the question, not the feature list. Teams under European data-protection rules, or any regime that limits cross-border transfer, cannot route personal data through a US-only session with no deletion guarantee. Our reading: run the Agents API where the data is boring, keep the Agents SDK or your own harness where the data is sensitive, and revisit once regional residency and retention controls arrive. Outsourcing the plumbing is tempting. The question of where the data lives stays yours.

Sources: OpenAI, Introducing the Agents API, OpenAI Developers, Agents API Overview, The Decoder