OpenAI runs your agent loop now. The fine print is US-only

On September 10, OpenAI put the Agents API into public beta. @stevendcoffey, who works on it, described it as "a brand new way to build Agents in the cloud, backed by the Codex harness" and told developers to point it at any sandbox they like. The official announcement went out from @OpenAIDevs the same minute.

If "harness" means nothing to you, it is the part of agent software that eats the engineering budget. A model that can reason is available to anyone with a credit card. The harness is everything around it: the loop that decides when to call a tool, what to do when that tool returns garbage, how to trim a conversation that has outgrown the context window without losing the thread, how to pick the work back up after the process dies at 3am. Teams have burned months on that code. OpenAI now runs the same harness that powers Codex, and you configure it instead of writing it.

What actually shipped

You define an agent (model, instructions, tools, MCP servers), optionally attach a sandbox environment, and start a durable session that streams events back to you. OpenAI handles orchestration, context compaction, and recovery. You keep your tool definitions and you choose where the code runs: OpenAI-hosted sandboxes, your own, or one of nine partners including Cloudflare, Modal, E2B, DigitalOcean, and Vercel.

There is no additional fee. You pay for tokens, tools, and container time, the same as before. A company that gives away its orchestration layer has decided that layer is now plumbing rather than something it can sell you.

The launch numbers came from early customers. SafetyKit reported cutting cost per case by 60% after moving its review workflow over. Hypha reported an 86% drop in failed agent responses, which it got by separating agent execution from the sandbox environment. Another early user reported a 4x latency improvement on subagent flows. Those are vendor-supplied figures on a launch page, so discount them, but the mechanism is believable. Less homemade plumbing means fewer places to break.

The line nobody posted

OpenAI's own documentation for the Agents API says it "currently supports data residency only in the United States and does not support Zero Data Retention." The docs add that picking a self-hosted sandbox does not change that.

Read that twice if you handle other people's records. OpenAI has publicly committed to keeping zero data retention available for frontier models, paired with a preview of something called Private Safety Processing. The newest and fastest way to build an agent on its platform does not qualify. If you operate under GDPR, or you have signed something promising a client that no vendor stores their data, the shortcut is closed to you this week.

It is a beta constraint and it will probably lift. Notice where it lands, though. The work that is most obviously worth handing to an agent is repetitive and rules-heavy, which describes healthcare, legal, insurance, and finance, and those are the sectors that cannot take the deal yet.

Two bets on the same day

OpenAI also shipped GPT-Live-1 to the API on September 10, a full-duplex voice model that @OpenAIDevs said works with "the models and harness you choose." One product asks you to hand over the loop, the other assumes you brought your own. My read is that OpenAI has not settled the question of whether owning the harness is worth defending, and is covering both answers.

That hedge is useful to anyone building. The loop is becoming a commodity, but the written description of how your business actually handles a quote, a claim, or an intake is not, and neither are the tool definitions that reach into your systems. Keep those in your own repository as plain files and swapping runtimes is an afternoon of work rather than a rebuild.

What this means if you run a business

The cost of standing up an agent that runs for hours and recovers from its own mistakes dropped sharply this week. The cost of knowing which of your processes deserves one did not move at all, and that is the number that decides whether any of this pays for you.

Most owners we talk to are still guessing. They know the front desk is drowning and the monthly reporting eats a Saturday, but nobody has timed the steps. If you want that mapped before you spend anything, New Face Design does a free process audit: we walk your actual workflow, put hours and dollars against each handoff, and say which pieces are worth automating. Sometimes the answer is leave it alone, which is a real answer and costs you nothing to hear.