Selfship helps teams running AI agents turn repeated customer-facing failures into reviewable fixes. It observes live conversations, finds patterns in what went wrong, proposes an eval-backed GitHub pull request, and checks whether the merged change holds on new traffic.

What's included in this release:

• Live conversation observation gives teams a grounded view of what customers tried to do and whether the agent helped.
• Failure detection groups repeated problems so engineers can focus on issues affecting many customers instead of isolated tickets.
• Eval-backed pull requests bring proposed fixes to GitHub with impact, sample sessions, and scores that reviewers can audit.
• Human approval controls keep every change reviewable, with path restrictions and limits on how often fixes are proposed.
• Post-merge verification watches real traffic and reopens a fix path when the improvement does not hold.

Together, these capabilities connect observation, diagnosis, repair, review, and live proof into one continuous improvement loop for production agents.