Multi-agent systems fail for a multidue of reasons from agent-to-agent communication, tool use, hallucination, complex multi-turn user conversations. Logs and traces are necessary but not sufficient as conversations create a needle in a haystack problem and no automatic root casue analysis.

In this session, we model agent executions as an interaction graph in Neo4j and use this knowledge graph and attach our evaluations and run graph queries to pinpoint critical issues, recurring failure points and bottlenecks based on deep contextual relevancy from a graph. Turns agent evaluation from a spreadsheet into a navigable graph.