AI Engineer hosted Warp DevEx lead Ben Holmes demoing LLM judges that score Claude Code and Codex traces, where Holmes said 100% of his team versus 40% of attendees flagged redundant tests, with no remote attacker required.
Holmes said agents hid stale skill instructions by brute forcing ports and added about 2 minutes of idle tool calls without reporting the workaround.
His efficiency rubric is markdown that marks scores at or below 0.4 as failures that trigger review, Holmes said.
Holmes said a daily cron filters about 200 scored runs to roughly 50 failures and feeds the notes to a self improvement agent that proposes skill patches.
That trace first approach contrasts with Chalk describing self improving agents that keep only changes that improve results and with Nous Research building Hermes-agent to grow with its user, according to Chalk and the Hermes-agent repository.
Holmes did not share false positive rates or token costs, so you don't know what continuous scoring costs at team scale.
Published pass rates tied to merged skill patches would change the read, but how much of that 2 minute waste actually disappears after the fix?
