Agent session waiting on a cold sandbox? Pre-warm it while the user types

#What changed

OpenAI's Agents API now lets you create and prepare an OpenAI-hosted sandbox — install packages, load files, run setup commands — before you open an agent session. When the task arrives, attach the prepared environment_id so the user is not waiting on a cold-boot setup.

Kevin Whinnery (OpenAI developer experience) announced the pre-warming flow on Oct 12, 2026 (source). Official docs: https://developers.openai.com/api/docs/guides/agents-api/environments/prewarmed

#Who is affected

Developers building on the Agents API where an end user types a request or picks a dataset before the agent needs a sandbox — for example a support agent or a data-analysis agent.

#What to do now

1. When the user starts interacting (typing in a field, choosing a dataset), create an environment with POST /v1/agents/environments, including the packages, files, and setup_commands your task needs.
2. Wait until the environment status is ready (poll retrieve, or use agent.environment.ready / failed webhooks). You can also attach while setup is still running and wait on session environment.ready events.
3. Create the session and pass that environment_id under environment.type openai_hosted.
4. Pre-warm close to the task: unattached pre-warmed sandboxes expire after about 5 minutes. Pre-warming is free until you attach; after attach, standard container pricing applies. After a session uses it, idle expiry is about 20 minutes of inactivity.

#What is not confirmed

Whether the Codex CLI desktop/interactive daemon uses this same Agents API pre-warm path. The docs cover Agents API hosted environments, not a CLI-side pre-warm switch.