OpenAI introduced the Agents API in public beta on September 10, giving developers access to the managed harness and infrastructure that powers Codex. The API is built for long-running cloud agents that can manage context, use tools, coordinate subagents, work with files, run code, and save intermediate results.
The important details are operational. OpenAI says agents can run in OpenAI-hosted sandboxes, self-hosted infrastructure, partner sandboxes, or customer VPC deployments. The first sandbox partner list includes Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. During beta, OpenAI says there is no additional Agents API fee beyond tokens and tools used.
Grey Haven’s read: this is a signal that agent infrastructure is hardening into a separate buying and architecture layer. The model matters, but the harness now owns context windows, file state, tool search, secrets, environment setup, subagent fanout, and long-session recovery. That is production surface area, not glue code.
Operators should treat managed agent runtimes the way they treat workflow engines or CI infrastructure. Ask where code runs, how secrets are injected, what the network can reach, how intermediate artifacts persist, what logs exist, and who can kill or resume a job.
The watch item is cost and authority drift. If agents can run for hours or days, the control plane needs budget limits, permission boundaries, and review gates before broad deployment. Source: OpenAI, “Introducing the Agents API,” September 10, 2026.
