Do LLM Agents Negotiate Rationally? A Mechanism-Design Framework for Verifiable Multi-Agent Interaction over A2A/MCP

Details

As multi-agent AI systems interoperate through standards like Anthropic's Model Context Protocol (MCP) and Google's Agent2Agent (A2A) protocol, a fundamental question emerges: do LLM agents behave rationally in multi-party interactions? Albayaydh and Zhao apply classical mechanism design theory to answer this empirically. The framework encodes negotiation mechanisms — alternating-offers bargaining and VCG-style auctions — as constraints over A2A message schemas, adds a runtime verification layer checking protocol invariants, and evaluates multiple LLM backbones (N=30 per condition). Across negotiation trials: unstructured dialogue achieves approximately 97% and 93.3% success across two models; structured protocols with verification achieve 100% for both. The auction results are more revealing. Both models achieve 100% efficient allocation, but truthful bidding diverges sharply: one model bids its exact valuation in every trial; the other does so in only 3.3% of trials. This is the key finding — mechanism-level incentive compatibility does not automatically transfer to LLM-agent behavior. Classical mechanism design cannot simply be imposed on LLM agents and expected to produce rational strategic behavior. A three-party fair-allocation task produced only 4.2% usable outcomes, reported as a negative result. The paper explicitly bridges classical multi-agent systems theory with modern LLM-agent infrastructure.