Warp's Ben Holmes: Agents That Review Their Own History and Open PRs to Fix Themselves
AI Engineer · youtube · 2026-11-07
Warp's agent self-improvement loop
Ben Holmes of Warp demos how to make AI coding agents improve their own setup:
- Collect everything: gather all agent conversations (with tool calls and outputs) in one place for analysis.
- LLM-as-judge scoring: rubric-based scoring for redundant tests, verbosity, and efficiency, including a live grading of a real conversation.
- Daily fixer agent: an agent that finds failure patterns daily and opens pull requests to fix skills and config.
Key insight
Agents quietly work around broken instructions without telling you — which is exactly why systematic review of agent history matters.
He also demos Skill Doctor, a skill you can run on your own Claude Code or Codex history, and shows how to track scores over time.
More from coding & agent
- Claude auto-renders 21k Blender images to train a model that recreates your LEGO build — nptacek · 2026-11-07
- The anti-slop AI coding prompt that works: "deepen modules, simplify interfaces" — mattpocockuk · 2026-11-07
- AI coding widens the engineer gap: average users vs power users is night and day — bendee983 · 2026-11-07
- Matt Pocock: to constrain coding agents to happy paths, you need to know the map — mattpocockuk · 2026-11-07
- Survey of 520 tech leaders: only 8% use one model provider, 80%+ delayed migrations over switching costs — ombre-ai · 2026-11-07
- Claude paints watercolor trees via custom-SDK abstractions, not direct robot control — teortaxesTex · 2026-11-07