Autonomous agents are starting to do business with one another, often representing different principals — a buyer's personal agent haggling with an enterprise's seller agent, for example. But when these agents keep their internal state private, things can go wrong in ways that are hard to diagnose.

A new preprint on arXiv proposes a failure taxonomy for exactly these situations. The authors examine agent-to-agent systems in which each agent acts for a different person or organization and does not share its full internal state. They categorize the kinds of failures that emerge from these boundary-crossing interactions, offering a structured way to identify why a multi-agent system broke down.

The taxonomy is significant because it moves beyond debugging individual agents and focuses on the interaction layer itself. By naming the failure modes, the paper gives developers a common vocabulary for designing more robust cross-principal agent systems.