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 predictively when the user begins typing, not when they submit the prompt.

Set git config user.name and user.email explicitly in the sandbox for each prompt, and create PRs using the prompting user's GitHub OAuth token rather than the app identity. Use GitHub App installation tokens only for repository cloning during image builds.

Yes, allow file reads immediately while blocking only writes until synchronization finishes. In large repositories, prompts rarely touch files changed in the last 30 minutes, so read-time staleness is safe while write conflicts are not.

The reference implementation uses Modal sandboxes with filesystem snapshots and near-instant startup. Cloudflare Durable Objects handle per-session state with WebSocket hibernation, and 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, 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 and revision rates, and the percentage of agent-written code across repositories.