OpenAI turns Codex's agent loop into an API

OpenAI opened its Agents API to every developer on September 10, turning the machinery behind Codex into a paid service.

That service runs on automatic memory compaction. A separate Tencent study found that when compaction dropped an agent's control constraints, agents crossed authorization boundaries in 87% of runs, but when the constraints survived, the rate was 0%.

It replaces the sandbox code and memory management most teams have been writing by hand on top of the Responses API. There's no separate platform fee, just tokens, tools and container time.

The system has four pieces: an Agent, an optional sandbox Environment, a durable Session, and streamed events. Sandboxes run on OpenAI's own hosts, on a developer's machines through codex exec-server, or through nine partners, including Cloudflare, Modal, DigitalOcean and Vercel. Subagents get their own context, and the service compacts that context automatically once it fills up.

Data stays in the US, and Zero Data Retention isn't an option, ruling out regulated workloads and any contract with retention clauses.

The overview doesn't describe what compaction keeps, or offer a way to pin instructions so they survive a summary, per OpenAI's Agents API overview.

Six months ago, a solo developer building this would have reached for LangGraph or written the agent loop by hand. Now a vendor sells the loop, the sandbox and the memory management, billed by usage, the way most teams stopped writing their own auth.