AgentAudit: Visual Analytics for Tracing Behavioral Escalation in Multi-Agent Systems

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.