Let Agents Rewrite Their Own Harness — But Keep the Grader Independent

DESI_MOGGER · reddit · 2026-09-12

A Reddit proposal for testing genuine agent self-improvement: let the agent rewrite parts of its own harness (prompts, tool handling, retries) while keeping the grader completely outside the loop. Running against the same independent evaluation after each change makes improvement measurable and prevents the agent from quietly optimizing its own scoring logic instead of actually getting better.

Original post →

More from coding & agent

coding & agent channel →