hosted-agents

Pre-build environment images on a 30-minute cadence so sandbox startup only requires a fast git delta sync. Maintain warm pools of ready sandboxes and start warming predictively when the user begins typing, not on submit.

Set git config user.name and user.email explicitly in the sandbox for each prompt, since background agents are not tied to a user during image builds. Create PRs through the API layer using the prompting user's GitHub OAuth token, not the app identity.

The reference implementation uses Modal sandboxes for near-instant startup and filesystem snapshots, with Cloudflare Durable Objects for per-session state and WebSocket hibernation. The patterns adapt to other providers like Fly.io with equivalent isolation models.

Yes, file reads are safe before sync completes because prompts in large repositories rarely touch files changed in the last 30 minutes. Block only writes until synchronization finishes to prevent conflicts, queueing pending writes in the meantime.

Work is lost when sandboxes are recycled or time out before results are extracted. Always snapshot the filesystem before termination and build explicit extraction steps, such as pushing branches and creating PRs, into the session teardown flow.

Track sessions resulting in merged PRs as the primary success metric, along with time from session start to first model token, PR approval rate, and agent-written code percentage per repository. Avoid vanity metrics like raw session counts.