We built a framework that analyzes LLM agent behavior during execution instead of only final outcomes.
It detects failure signals such as:
- tool errors
- repetitive actions
- stagnation
- degraded exploration
This helps understand how and when LLM agents begin to fail during runtime, not just after completion.
