"Agentic societies need a social harness": honest, competent agents still fail to coordinate across trust boundaries with today's messaging primitives

arXiv·medium signalXBlueskyLinkedInCopy link

University of Washington researchers argue the harness concept needs splitting in two — the personal harness each agent already has for private context and principal communication, and a missing social harness governing inter-agent interaction across trust boundaries where principals' objectives only partially align. They show experimentally that even honest, competent agents often fail to reach satisfactory outcomes with existing harnesses, and that faulty or malicious agents can stall collaboration, swing outcomes and pursue other goals purely by exploiting vulnerabilities in agent speech. Their proposed layered architecture prevents some failure classes outright, lets agents detect invalid messages at runtime, and supports post-facto investigation and consequences — a useful frame if you are wiring A2A or MCP across organizational boundaries.