OpenAI’s Agents API puts the Codex harness in a public beta
Developers get durable sessions, context compaction and hosted or bring-your-own sandboxes. There is no extra subscription: tokens and tools only. The changelog dated the beta to September 10.
Short1 minEasy
Priya Shah1 min read
- OpenAI
- Agents API
- Codex
- Developers

Photo: OpenAISource page
OpenAI’s API changelog lists a public beta of the Agents API on September 10. The company says it is the same managed Codex harness it already uses internally, exposed as sessions that developers can start with a single call. OpenAI handles orchestration, context compaction and recovery. You choose the model, instructions, tools and where code runs.
Sessions can last across turns and stream progress. Agents connect to your tools and MCP servers. Environments are either OpenAI-hosted sandboxes — code, files, artifacts — or a sandbox you bring. The community post names first-party integrations with Blaxel AI, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle Cloud, Runloop AI and Vercel. Requests take an `OpenAI-Beta: agents=v1` header; the official SDKs add it. The documented example uses `gpt-6-astra` in an `openai_hosted` environment.
There is no extra Agents API subscription. Developers pay for tokens and tools. Multi-agent mode, when enabled, lets a coordinator spawn subagents with a default cap of six concurrent children. The beta is not Codex-the-product; it is the loop behind Codex, now on the same key as Responses. OpenAI says it will iterate toward general availability from feedback.


