OpenAI opens Codex agent harness to all developers

What happened: OpenAI launched the Agents API in public beta at 8am Pacific on 10 September 2026, the same morning it released GPT-Live-1 and ChatGPT for Financial Services.

The details: The API exposes the harness that already powers Codex and ChatGPT for Work: sessions, orchestration, automatic context compaction, tool search, subagents and recovery. One call sets the task, model, tools and environment. Four building blocks make it up: an Agent, an optional sandbox Environment, a durable Session, and streamed events.

Who's affected: It's open to all API developers, not just enterprise customers. Agents can run for hours or days, with sandboxes hosted on OpenAI's own infrastructure, self-hosted in a customer's cloud, or run via partners including Cloudflare, DigitalOcean, Oracle and Vercel. OpenAI always runs the harness itself.

The numbers: OpenAI said: "There are no additional fees for using the Agents API." Customers pay for model tokens, tools, and OpenAI-hosted container time only.

The catch: Data storage is US-only for now, and Zero Data Retention isn't supported yet.

Who's using it: Early customers named include Ciridae, Long Lake, WithCoverage, SafetyKit, Dwelly, Hypha, deepsense.ai and Nash.ai.

Why it matters: The scaffolding around a model increasingly determines what an agent can actually do. By productising its own harness, OpenAI is competing directly with Anthropic's Claude agent tooling for developers building long-running, tool-using agents.

What's next: As a public beta, the API is likely to expand in coverage, including eventual support for data retention controls beyond the current US-only storage limit. OpenAI's announcement describes the Agents API as "fully managed by OpenAI."