Built this after spending too many hours manually debugging why our AI agent kept failing on real world examples.

evalfix runs your evals, identifies what's failing, and uses a multi-agent loop to diagnose and fix your prompt — then verifies the fix doesn't break anything that was passing.

It also ships with an SDK to capture production failures and automatically pull them into your eval suite.