Building feedback loops that make your AI coding agent smarter from its own mistakes, one correction at a time.
3 parts
Two unrelated projects discovered the same pattern: point an AI agent at code, give it a score to chase, and let it run experiments until morning. I generalized it into a tool that works for any domain.
Your AI coding agent makes the same mistakes over and over. What if it could learn from corrections, track which skills cause failures, and tell you whether it already fixed the problem? I built a closed-loop learning system for my coding agent, inspired by a meta-learning paper, and here's how it works.
Anthropic charges $0.08/hr for rubric-based agent evaluation. Here is how to build the same pattern the open-source way.
