"Agent 协议栈三层分化：MCP 管工具、A2A 管协作、Agent Plugins 管分发"
Agent 协议栈三层分化：MCP 管工具、A2A 管协作、Agent Plugins 管分发 2026年8月，协议战结束了。 这不是说竞争消失了——而是战场发生了转移。半年前，开发者还在热烈争论"MCP vs 其它协议哪个更好"，而现在答案已经清晰得让人有点意外： 没有哪个更好，只有哪个更适合哪一层。 三层架构是如何成形的 先说背景。AI Agent 的协议演进，在过去两年里经历了一个典型的"分久必合、合久必分"的过程： 2024年 ：各自为战，每家平台都有自己的工具调用格式 2025年 ：MCP 异军突起，成为工具调用的事实标准 2026年上半年 ：协作需求爆发，A2A 协议顺势而出 2026年8月 ：打包分发问题浮出水面，Agent Plugins 1.0 登场 三层架构不是设计出来的，是需求倒逼出来的。 第一层：MCP —— 管"工具"，做执行 MCP（Model Context…
The AI Agent protocol stack has undergone a three-layer division: MCP as a tool manager, A2A for collaboration, and Agent Plugins for distribution. This division is a result of demand, rather than a design choice. MCP, the first layer, focuses on agent-to-tool interface, with no state and HTTP architecture. It has become the infrastructure for the entire AI tool ecosystem, with over 110 million downloads.
A2A, the second layer, deals with agent-to-agent collaboration, offering a clean interface for communication, context exchange, and task coordination. A2A has gained significant traction, with adoption by major cloud platforms. Agent Plugins, the third layer, focuses on packaging and distributing agent skills, allowing for cross-platform and environment deployment.
Anthropic's absence from the Agent Plugins' joint release signifies a conscious decision to maintain protocol neutrality. The next battlegrounds are packaging distribution, trust and security, and persistent memory. The emphasis is on developing modular, composable skills rather than monolithic agents. The future of AI agents will revolve around their ability to be packaged, distributed, and collaborated on as digital labor units.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
