See which agent broke your run.

AgentTrace draws a multi-agent run as one line, marks where the mistake came in, and lets you fork any step to test a fix before you ship it.

A real saved run. Select any node to see what it is.

Original run

Fork 1: critic changed

Fork 2: planner changed

Select a node on the line to see what it is, what it received and what it said.

How you use it

Record
Add the Python recorder to your agents. Each run becomes a line of agents, with secrets removed first.
Find the cause
Open a run that went wrong. The agent that most likely caused it is marked on the line.
Fork to test a fix
Change what one agent decided. A new line opens and the rest runs again, with and without your change.