Skip to content- Sources: primary
- Summary: Ben Holmes describes collecting each agent session with its transcript, tool calls and resulting diff in one place, then grading it with markdown LLM-as-judge rubrics that set pass and fail thresholds. A daily cron agent reads the failing reports, finds common causes and opens pull requests against skills and agent config. Examples include a test skill that wasted tool calls hunting for free ports and a triage skill missing config to spawn subagents. Warp released a one-off version as a skill called Skill Doctor.
- Why it matters: It gives a concrete loop for finding silent tool-call waste in unattended agents and measuring whether a skill change helped.