CAIN-42 CAIN StudioCAAE · Module 5 of 5 · 11 min
Root-causing multi-agent failures
When a multi-agent run fails, the visible error is often far from the cause: a sub-agent's bad summary three steps earlier.
A root-cause tracer walks the causal chain of messages and actions back to the first step whose output violated its contract.
Store causal links (which step consumed which output) at run time; you cannot reconstruct them reliably afterwards.
Key points
- Visible error ≠ cause.
- Walk causal links back to the first contract violation.
- Record causal links at run time.
In the CAIN-42 codebase
multi-agent-failure-tracer/— root-cause attribution for broken multi-agent runs
Practice check
Practice 1 · choose one
The root cause of a multi-agent failure is…
Practice 2 · true or false