Multi-agent self-improvement loop: scorer agents grade past runs, then auto-PR AGENTS.md fixes
blaizedsouza · x · 2026-09-23
Ben Holmes outlines a practical multi-agent "self improvement" system where agents review their own past failures and codify the lessons:
- Scorer agents: review past conversations in batches and assign pass/fail grades along metrics you define (code quality, efficiency, task compliance). Run as a subagent swarm, or as an automation using hooks in your agent harness.
- Self-improvement agents: receive failing transcripts plus the scorer's report, analyze what went wrong, and open a PR updating skills or AGENTS.md files — citing the failed runs as evidence for each change.
The result is an automated learning loop that turns past failures into persistent agent behavior updates, directly adoptable by teams already managing agents via skills/AGENTS.md.
More from coding & agent
- Agentic DORA metrics: tracking PR start-to-merge to measure AI coding agents — vincent_koc · 2026-09-23
- Replacing Claude & Chrome with Strawberry, an AI browser with built-in agents — damienghader · 2026-09-23
- Claude Opus 5.5 impresses with TTFT and throughput in Claude Code daily use — rachittshah · 2026-09-23
- DHH: Over 4,000 Omarchy Plugins Published as the Agentic OS Ecosystem Takes Off — AIFlow_ML · 2026-09-23
- AI agents in a kingdom game learned to starve their own people to push refugees at neighbors — Typical-Pollution327 · 2026-09-23
- Dev finds Claude Code smoother at reviewing PRs than at writing code — JasonBotterill · 2026-09-23