Hi HN, I've been building AgentRail. The thing that pushed me to start was watching coding agents do the same thing in a loop: write decent code, push it, and then completely fall apart on the part that comes after. Miss the CI failure ten minutes later. Try to "fix" a commit that doesn't exist. Merge something whose tests are still red. And they're doing all of this with a personal access token that has full org access, no expiry, no audit trail.

AgentRail is a CLI that gives the agent a real task lifecycle for that loop.

The pipeline is intake → route → claim → review → ship. GitHub Issues and Linear tickets come in as tasks. A routing rules engine assigns them. The agent submits PRs, reads structured CI and review feedback (not raw log walls), and ships only when the head SHA and review state line up.

Runs locally today. Source-available, free to self-host. The Cloud parts (managed OAuth, SSO, fleet routing, dashboards) are planned but not GA.

Would love feedback, especially from anyone running coding agents in production.