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. Start warming a sandbox as soon as the user begins typing so it is ready before they submit their prompt.

Set git config user.name and user.email explicitly in each sandbox using the prompting user's identity. Use GitHub App installation tokens for cloning, but create PRs with the user's OAuth token so PRs appear as human-authored.

Yes, file reads are safe before sync completes because recently changed files rarely overlap with the agent's research targets. Block only write operations until synchronization finishes to prevent conflicts with incoming changes.

Modal Sandboxes provide near-instant startup and filesystem snapshots suitable for this architecture. Cloudflare Durable Objects handle per-session state with WebSocket hibernation, while the patterns adapt to other providers like Fly.io.

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 like pushing branches or 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 response, PR approval rate, and the percentage of agent-written code across repositories.