elvis@omarsar0 · X该报告针对LLM多智能体系统的通信瓶颈，构建了五维分类法（对方、有效载荷、交互状态、发现机制、模式灵活性），系统梳理了9个积极维护的开源智能体协议，覆盖MCP和A2A的实际格局。报告发现两个突出模式：每个智能体间协议都采用混合有效载荷与会话状态持久化组合，而去中心化发现机制仍极为罕见。领域正悄然标准化有状态会话，但发现与策略执行层仍留白。该报告为今年选择通信层时提供了九大协议的真实对比参考。
Great report on LLM agent communication protocols.
Communication is a huge bottleneck in multi-agent systems.
（worth bookmarking）
The report builds a five-dimensional taxonomy （counterparty， payload， interaction state， discovery mechanism， schema flexibility） across nine actively maintained open-source agent protocols， so it maps the real MCP and A2A landscape.
Two patterns stand out. Every agent-to-agent protocol sampled pairs of hybrid payloads with session-state persistence， and decentralized discovery is still rare. So the field is quietly standardizing on stateful sessions while leaving discovery and policy enforcement open.
Why does it matter？