Warp Factories that score themselves and open fix PRs
Warp’s self-improvement loop: scorers for efficiency/quality/compliance sample runs, then a scheduled agent proposes skill/config diffs from failures.
Standing up a factory connected to issues and forges is easy; making it learn your team is the hard part. Warp’s answer: file-defined scorers that classify agent runs (e.g. efficiency from highly inefficient → clean) with pass thresholds and sampling rates, then a scheduled self-improvement agent that batches failures and proposes skill/config changes as reviewable PRs.
They dogfood scorers for code quality, efficiency, and procedure compliance. Examples of merged self-improvement: stop 275 computer-use actions to verify one CSS line; add a Design dimension after agents warped shared abstractions; cut no-op orchestrator messages; fix broken skill references; heartbeat from Slack-dispatched agents.
Guidance against LLM-ism: redraft skills holistically, don’t pile addendums.
Why it matters for code review
Review quality in a factory is a living system. Scoring runs and feeding failures back into reviewer skills is how you keep agent comments high-signal instead of slowly rotting into noise.