An exploration of agent evaluation, from concepts and implementation to real experiments. After changing a coding agent's system prompt, does it actually get better at fixing code? I wanted a project ...