hosted-agents

Pre-build environment images on a 30-minute cadence with dependencies installed and caches warmed, then maintain a pool of pre-warmed sandboxes per repository. Start warming a sandbox as soon as the user begins typing so it is ready before they submit.

Use GitHub App installation tokens for cloning inside the sandbox, then set git config user.name and user.email to the prompting user's identity. The API layer uses the user's OAuth token to open the PR, so it appears authored by the human.

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

Yes, allow file reads immediately because recently changed files rarely overlap with the agent's research targets. Block only writes until synchronization finishes, queuing pending writes and flushing them once sync completes.

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, plus time from session start to first model response, PR approval rate, and agent-written code percentage per repository. These indicate real value rather than vanity usage counts.