OpenAI Agents API: What the Codex Harness Gives You

OpenAI Opens the Agents API to Every Developer

OpenAI put its Agents API into public beta on September 10, and the short version is that the company has taken the machinery it built for Codex and turned it into a service anyone can call. Until now, if you wanted an AI agent that ran for an hour, survived a crashed step, kept its context tidy and coordinated a handful of subagents, you wrote that scaffolding yourself. The Agents API moves that layer onto OpenAI's side of the wire.

• Availability: public beta for all developers as of September 10, 2026, with US-only data residency and no Zero Data Retention support yet
• Pricing: no separate fee for the API — you pay for tokens, tools and container time
• Execution: OpenAI-hosted sandboxes, your own infrastructure via the Codex exec server, or one of nine partner sandboxes
• Partners at launch: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel

What Does the Agents API Actually Manage?

The API is organized around four ideas, and they are worth separating because each replaces a piece of code most teams have already written twice. An agent is a model plus instructions plus tools. An environment is an optional sandbox where that agent can run commands and touch files. A session is a durable instance of an agent working on something. Events and items are the inputs and outputs streaming in and out of that session.

Durability is the part that changes the engineering calculus. In the usual pattern, an agent's state lives in your process, so a long task is a long-held connection and a failure is a restart from zero. A session here persists across turns and tasks, and progress streams back to your application while the work continues. OpenAI also handles context compaction automatically as a session approaches its limit, which removes one of the fiddliest pieces of agent plumbing — deciding what to summarize, what to drop, and when.

Tool handling gets a similar treatment. The API supports Model Context Protocol servers, custom functions and built-in tools such as web search, and it uses tool search to load tool definitions only when they are needed. That matters more than it sounds: a large tool catalog inflates every prompt, and loading definitions lazily keeps the token bill down while preserving cache efficiency.

Why the Sandbox Partner List Is the Interesting Part

Most agent frameworks assume you supply the execution environment. Naming nine sandbox providers with first-class integrations at launch is a statement about where OpenAI expects the work to happen — near the developer's existing infrastructure rather than inside a single hosted box. If your code and data already sit on Cloudflare, DigitalOcean or Oracle Cloud, the agent can run beside them instead of reaching across the internet for every file read.

The self-hosted route is the other half of that story. Running the Codex exec server on your own machines means the harness is managed by OpenAI while the execution is not, which is the split a lot of regulated teams have been asking for. The residency caveat is real, though: US-only data residency and no Zero Data Retention at beta will rule the service out for some organizations until those land.

Do the Early Results Justify the Switch?

OpenAI's announcement cites early customer figures — evaluation scores moving from 0.71 to 0.85, cost reductions near 60%, and 86% fewer failed responses. These are the company's own numbers drawn from selected customers, and no independent benchmark of the Agents API has been published yet, so read them as directional rather than settled.

The structural argument is stronger than any single figure. Context compaction, session recovery and subagent coordination are exactly the pieces that are tedious to build, easy to get subtly wrong, and rarely a source of competitive advantage. Handing them to a managed runtime is the same trade teams made when they stopped running their own queues. What you give up is control over how compaction behaves — and for agents doing precise, long-horizon work, that policy is not a detail.

For context on how quickly this stack is assembling, the Agents API landed the same day as the GPT-Live-1 voice API, which is designed to hand its heavy reasoning to exactly this kind of backend agent. OpenAI's own coding-agent velocity report covers what these harnesses are doing internally. More releases are in our AI news coverage.

Sources: OpenAI — September 10, 2026; MarkTechPost — September 10, 2026.