I recently made halftrace - point it at your OpenAI/Anthropic/LangSmith agent logs and it tells you what shape your agent's rule-following has, why, and what to try. I built it to try and find smooth decay over trajectory length on Claude.

https://halftrace.dev