OpenAI Agents API public beta: a developer's screen showing code for creating AI agents with a single call.

Editorial illustration for OpenAI's Agents API Enters Public Beta, Creating Agents With One Call

OpenAI's Agents API Enters Public Beta, Creating Agents...

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

September 11, 2026 • 3 min read

OpenAI opened public beta access to its Agents API on Thursday, handing outside developers the same harness that already powers Codex and ChatGPT for Work. Until now, that infrastructure sat behind OpenAI's own products. A single API call can spin up an agent with a model, a set of tools, and a sandboxed environment to work in, then keep it running for as long as the task takes, days if necessary, while OpenAI hosts and maintains the underlying harness.

The pitch is straightforward: building an agent that can hold context over long stretches, call tools without wasting cycles, and hand off work to subagents is hard, and OpenAI says its own scaling work on Codex exposed exactly what that infrastructure needs to look like. Rather than making every developer solve that problem separately, the Agents API packages it as a managed service, with compute running either in an OpenAI sandbox, a developer's own infrastructure, or a partner's.

There are real constraints attached. Data stays US-only for now, and Zero Data Retention isn't supported, details that matter for anyone weighing this against a self-built agent stack. OpenAI's own documentation lays out the core building blocks developers now have to work with.

The Agents API is a managed service built on the open-source Codex harness. OpenAI team states scaling Codex and ChatGPT for Work showed what long-running agents need.

OpenAI Launches the Agents API in Public Beta, Putting the Codex Harness Behind One API Call - MarkTechPost

Why this matters

The Agents API strips out the part of agent-building that eats the most engineering time: standing up a harness that manages sessions, tool calls, and sandboxed execution. OpenAI is betting that developers would rather pay for that plumbing than build it themselves, and given how many teams have already reverse-engineered Codex-style patterns, that bet looks reasonable. But the US-only data residency and lack of Zero Data Retention support are real constraints, not footnotes. Any team in Europe, or any regulated industry handling customer data, needs to read that as a current blocker, not a detail to revisit later.

The three deployment options (OpenAI-managed sandbox, your own infrastructure, partner sandbox) suggest OpenAI knows enterprises won't hand over full control immediately. Founders should watch pricing once the beta tag comes off, since managed harnesses tend to get expensive at scale. Researchers building agent architectures now have a production-grade reference implementation to test against, which is worth more than the API itself. Watch whether Zero Data Retention support arrives before or after general availability.