MPAC: A Multi-Principal Agent Coordination Protocol for Interoperable Multi-Agent Collaboration

A team of researchers has introduced MPAC (Multi-Principal Agent Coordination Protocol), a new application-layer standard designed to solve a critical gap in the AI agent ecosystem. Current protocols like MCP (Model Context Protocol) and A2A (Agent-to-Agent) assume a single controlling principal, breaking down when agents from different people or organizations need to collaborate on shared tasks—like multiple engineers' coding agents editing the same repository or agents from different companies negotiating a deal. MPAC fills this void with explicit coordination semantics across five layers: Session, Intent, Operation, Conflict, and Governance, making intent declaration a precondition for action and representing conflicts as structured objects.

The protocol's specification is comprehensive, defining 21 message types, three state machines, Lamport-clock causal watermarking for ordering events, and optimistic concurrency control for shared state. The researchers have released open-source reference implementations in Python and TypeScript, backed by 223 tests and seven live demos. In a controlled benchmark involving a three-agent code review task, MPAC demonstrated a 95% reduction in coordination overhead and a 4.8x wall-clock speedup compared to a serialized, human-mediated baseline. Crucially, this speedup comes from eliminating coordination waits, not from compressing the AI models' own reasoning time, preserving the quality of individual agent decisions. The pluggable governance layer also allows for human-in-the-loop arbitration when needed, making the system practical for real-world deployment.