Definition
Plain language
A system that reads an AI agent's failure logs, figures out which part of its surrounding software is broken, and writes a code patch to fix it.
As stated in the literature
An offline harness-repair pipeline (abstract, diagnose, repair, validate) that localizes agent failures to specific harness layers, applies fixes from a curated catalog of repair operators distilled from real repo histories, and accepts patches under a regression-aware criterion.
Why it matters: It automates the tedious work of repairing the software scaffolding around an agent, which is often the real cause of failures rather than the model itself.
For example, after an agent keeps failing because its file-reading tool chokes on large inputs, HarnessFix would spot that, write a patch for the tool, and check the fix doesn't break anything else.
Heard on the show
“HarnessFix is the only method that systematically gets into the lifecycle, the observability, the verification, and the governance layers.”Episode 121 — When the Agent Says It's Done But Nothing Happened: Debugging the Harness, Not the Model