AgentAudit: Visual Analytics for Tracing Behavioral Escalation in Multi-Agent Systems
Luis Sante (Fundação Getulio Vargas), Dany Mauro Diaz Espino (Fundação Getulio Vargas), Leighton Estrada-Rayme (Fundação Getulio Vargas), Juanpablo Heredia (Fundação Getulio Vargas), Joel Perca (Fundação Getulio Vargas), Felipe Moreno-Vera (FGV), João Rulff (New York University), Jorge Poco (Fundação Getulio Vargas)
Jump to Discussion
Links
Sign in to access the Paper PDF.
Abstract
When corporate communication is delegated to autonomous agents, failures are rarely a single decision but a trajectory of small deviations that compound until a policy is broken. We present AgentAudit, a visual analytics system that reconstructs such escalation from agents' public posts, private deliberations, and internal reasoning, coupling topic evolution, event extraction, per-agent channel usage, and think-say dissonance, how far a public message departs from the agent's internal state. On the VAST Challenge 2026 data, AgentAudit traces the embargo breach to a gradual migration of coordination into private and anonymous channels, foreshadowed by dissonance from the earliest rounds and by an earlier episode that was monitored but never corrected.