What's this about?
An open-source toolkit for LLM-agent failure observability: the step where an agent visibly fails is often far downstream of the step that caused it, and existing trace-replay tools don't help attribute blame.
What exactly is "failure attribution"?
Finding the true root cause in a long chain of agent decisions — like debugging why a relay race was lost when the stumble happened three handoffs before the finish.
Why is it worth noting?
21↑ — agent debugging is this year's developer pain; tooling that assigns blame (and suggests recovery) is the missing layer under every agent framework.