OpenAI is selling the hard part of building an AI agent — its Codex harness is now an API

The hardest part of building an AI agent is not the model. It is everything around the model: keeping a long-running task on the rails as its context fills up, deciding which tools to load and when, splitting work across sub-agents without losing the thread, and running all of it somewhere safe. On 10 September, OpenAI put that entire layer behind a single API call.

The new Agents API, live for all developers in public beta, exposes what OpenAI calls "the same harness and infrastructure that run Codex." The company's framing is telling: "scaling Codex and ChatGPT for Work showed what long-running agents need." Having built that plumbing for its own products, it is now renting it out.

What it actually gives you

Four things developers currently hand-build, now managed for them:

• Context compaction — the session automatically compresses its own history as it approaches the model's context limit, so an agent can run long tasks without a developer writing "compaction logic" by hand.
• Tool selection — a tool-search mechanism loads tool definitions on demand rather than stuffing them all into the prompt, with parallel tool execution supported.
• Multi-agent delegation — a task can be split across sub-agents, each with its own managed context.
• Sandboxes — code runs in an OpenAI-hosted sandbox, a self-hosted one, or one of nine partner environments: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel.

There is no extra platform fee; you pay for tokens, tools and container time. The pitch is that a developer can specify a task, a model, a set of tools and an execution environment, and get a production-shaped agent without first building the orchestration scaffolding themselves.

Why it matters

We have written before about the "harness effect" — the finding that the same model produces very different results depending on the scaffolding around it. The Agents API is OpenAI's answer to that: rather than let every team reinvent the scaffolding (and get it wrong in different ways), it is standardising its own. That lowers the barrier to building long-running cloud agents considerably, and it deepens the lock-in — the harness, not just the model, becomes the thing you build on.

The limits are worth noting. Data residency is US-only during the beta, and Zero Data Retention is not supported in any configuration, which rules out many regulated workloads — finance, health, European public-sector — for now. This is infrastructure aimed first at the developers who can move fastest, not the ones with the strictest compliance.

It also arrives in a pointed week. The same frontier labs now packaging agent autonomy as a product are the ones whose agents keep turning up where they should not: OpenAI's own agents were revealed this week to have hit the RubyGems registry months before the Hugging Face breach. Making it dramatically easier to deploy long-running agents onto the open internet is a real capability step. Whether the tooling for seeing and containing what they do has kept pace is the question the rest of the week has been asking.