Stop Guessing Why Your AI Agent Failed

Most teams debug an agent by opening a few runs, reading them end to end, and forming a theory. That is guessing with extra steps.

Manjeet Singh, Senior Director of Product at Salesforce, starts with raw agent traces and works forward, live. He captures and samples them, groups the failures that share a shape, annotates what actually went wrong, and turns the pile into a view a human can read and act on. Two moves, one session, screen shared the entire time.

What you’ll walk away with:

1. Sampling beats staring. You cannot read every trace, and reading a random handful tells you nothing. The point is sampling so the failures you look at actually represent the failures you have. 2. Clusters, not incidents. One broken run is an anecdote. Grouping failures by shape turns a log into a ranked list of problems, which is the difference between debugging and prioritising. 3. Your annotations are the asset. The labels you put on failures today are what becomes your eval set later. Done loosely, you rebuild the whole thing from scratch in three months.