For a while now, watching OpenAI's Codex build and fix its own code has felt a bit like watching someone else's dog do an impressive trick — cool, but not yours to use. As of September 10, that changed. OpenAI opened the Agents API in public beta, and the managed infrastructure behind Codex is now available to any developer with an API key.
The Harness, Unboxed
The Agents API centers on four building blocks: agents (a model plus instructions and tools), environments (optional sandboxes), sessions (durable running instances), and events (the inputs and outputs flowing through them). Developers can spin up a working agent — say, one built to investigate production incidents — with a single API call, and the platform handles the plumbing: automatic context compaction for long-running sessions, dynamic tool management to keep token costs sane, and native support for multi-agent task delegation.
Deployment is flexible too — agents can run in OpenAI-hosted sandboxes, self-hosted environments, or through nine partner sandbox providers including Cloudflare, DigitalOcean, Vercel, and Modal. Pricing is straightforward: no extra platform fee, just the usual costs of tokens, tool calls, and container time.
Why "It Just Handles Sessions" Is a Bigger Deal Than It Sounds
Anyone who's built an agent from scratch knows the unglamorous 80% of the work isn't the clever prompting — it's session state, context window management, retry logic, and making sure your multi-agent setup doesn't quietly forget what it was doing three tool calls ago. OpenAI just open-sourced (well, API-sourced) the boring-but-critical infrastructure it built to solve that problem internally.
The catch worth flagging before you get too excited: data residency is US-only right now, and Zero Data Retention isn't supported yet, which rules this out for plenty of regulated industries until that changes. For everyone else, this collapses months of orchestration engineering into a single API call.
The agent framework wars have a new contender, and it's the company whose model most of the others are built on top of anyway.
If you're a business wondering what AI agents could actually automate in your day-to-day workflows — not the sci-fi version, the practical one — that's a conversation worth having before your competitors have it first; let's talk through where it fits for you.
Source: MarkTechPost