On September 10, OpenAI opened the Agents API to every developer in public beta. The pitch is simple: the same orchestration harness that runs Codex — sessions, sandboxes, tool calling, subagents — is now available behind a single API call, with no fee on top of the tokens, tools and container time you already pay for.

That matters because the hard part of shipping an agent was never the model. It was everything around it: keeping a task alive for hours, giving it a safe place to run code, stopping it from drowning in its own context, and getting it to hand work off without losing track. OpenAI has been solving those problems for Codex since 2025. Now it is renting the solution out.

What you actually get

Strip away the launch language and the Agents API is a managed runtime for long-running agents. Its main parts:

• Sessions that keep going. A session holds the agent's state for as long as the task takes, and when it approaches the context limit the API compacts it automatically. You no longer write the summarise-and-restart logic yourself.
• Sandboxes, three ways. Run agents on OpenAI's own Codex infrastructure, self-host with the open-source codex exec-server over a WebSocket, or pick one of nine launch partners: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel.
• Files and code. Agents can edit files, execute code and save intermediate results inside the sandbox, the same way Codex does on a repository.
• Tools without the token tax. MCP servers, custom functions and built-ins such as web search all plug in. A tool-search feature loads definitions only when they are relevant, and programmatic tool calling lets the agent chain, parallelise and filter tool results instead of round-tripping each one through the model.
• Subagents. A parent agent can delegate parts of a task to subagents that run in parallel, each with its own context, then collect the results.

The code examples in OpenAI's announcement run on GPT-6-Astra, and the harness itself is open source on GitHub, so you can read exactly what the API is doing on your behalf.

Who is already running on it

OpenAI shipped the beta with customer numbers attached. All of them are vendor-supplied, none are independently benchmarked, but they show what the harness is being used for:

• Ciridae says its evaluation score rose from 0.71 to 0.85 and latency fell fourfold after moving to the Agents API.
• SafetyKit reports a 60% lower cost per case in its trust-and-safety review workflow, with better latency.
• Hypha cut failed agent responses by 86%.
• Nash.ai uses it to keep thousands of long-running agents alive across a global logistics network.

Long Lake, WithCoverage, Dwelly and deepsense.ai are also named as early users. The pattern across all of them is the same: work that runs for minutes or hours, touches files or external systems, and previously needed a custom orchestration layer to survive.

What it costs, and where it falls short

There is no platform fee. You pay standard model token rates, tool costs, and container time for the sandbox. That last item is the one to watch: an agent that idles in a hosted sandbox for hours is billing the whole time, and OpenAI has not published a per-container rate in the announcement itself.

Two limits will keep regulated companies away for now. Data residency is US-only, and there is no zero-data-retention option yet. If you are in finance, health or the EU public sector, the self-hosted sandbox path helps with where the code runs, but not with where the conversation is stored.

There is also a trade-off in control. Compared with OpenAI's Agents SDK or the raw Responses API, the Agents API is the lowest-effort option and the least flexible one: state lives in a saved session configuration on OpenAI's side, and context compaction happens on OpenAI's schedule, not yours. Teams that have already built their own orchestration will want to look hard at what they would give up.

Why this is the real story of the week

Two days before this launch, Meta put Muse in front of consumers: an agent that reads your inbox, holds your card and acts on its own. The Agents API is the other half of that shift. Muse is what an agent looks like to the person using it. The Agents API is what it looks like to the company building it, and OpenAI has just made that part something you rent by the hour rather than build over a year.

The competitive question is not whether Anthropic and Google follow with equivalent managed harnesses; both already run agents at scale internally and the pressure to expose them is now obvious. The question is where the value settles. If the harness becomes a commodity you get for free with the tokens, the differentiation moves back to the model, the tools an agent can reach, and the data it is allowed to touch. That is a much better position for OpenAI than competing on orchestration code that anyone can read on GitHub.

For developers, the practical advice is short. If you have a task that runs longer than a single request, needs a place to execute code, and you have been maintaining your own loop to keep it alive, try the beta this week. If you need EU data residency or zero retention, wait, and say so loudly enough that OpenAI hears it.