OpenAI opened the Agents API in public beta on September 10. The release exposes the same managed harness that powers Codex, so developers can spin up cloud agents with a single call while OpenAI handles orchestration, session state, and context compaction.
The model surface includes gpt-6-astra. Tools arrive through MCP servers or built-in capabilities. Execution environments range from OpenAI-hosted sandboxes to partner sandboxes (E2B, Modal, Vercel, DigitalOcean and others) or a developer’s own infrastructure. There is no separate fee for the Agents API itself; usage is billed on tokens and tools.
Key takeaways
- Agents API public beta launched September 10 on the Codex harness.
- Developers supply tools, model, and environment; OpenAI runs the agent loop.
- Hosted sandboxes and first-party partner integrations available at launch.
What the Agents API threshold actually means
Long-running agents have been hard to keep reliable because the harness—prompt formatting, tool routing, recovery, and context windows—had to be rebuilt for every new model generation. OpenAI is productizing that layer. The API supports multi-agent configurations, vault-backed secrets, and automatic context compression when sessions approach limits.
Design partners and early documentation emphasize internal automation and customer-facing agents rather than pure research demos. The release sits alongside the recent ChatGPT for Financial Services workspace and the broader Astra rollout. For the content-stack angle see How AI is Reshaping the Content Creator Stack in 2026.
What to watch next
Watch the first production case studies that use hosted sandboxes versus bring-your-own compute. Watch whether the public beta expands tool search and parallel tool calling beyond the initial surface.