A2A協定評測：併入MCP陣營，AI代理終於同語言 | A2A Protocol Review: Joins MCP at Linux Foundation

8月 28, 2026

By Kit 小克 | AI Tool Observer | 2026-08-29

🇹🇼 A2A協定評測：併入MCP陣營，AI代理終於同語言

AI代理開發最大痛點之一，是各家協定各吹各的調——這正在改變。Google的A2A協定（Agent2Agent）2026年8月正式併入Linux基金會旗下的Agentic AI Foundation（AAIF），跟Anthropic捐出的MCP成為同一治理傘下的專案，代理生態系從「各家搶標準」走向「共通協定」。

什麼是A2A協定？

A2A是讓不同廠商的AI代理彼此「發現」與「溝通」的開放標準。代理透過Agent Card描述能力與端點，讓其他代理能找到它、委派任務、交換訊息，不管背後是哪家公司做的。

A2A跟MCP差在哪？

簡單說：MCP負責「代理連工具」，A2A負責「代理連代理」。MCP解決模型怎麼安全呼叫資料庫、API；A2A解決多個代理間怎麼協作完成複雜任務。兩者現在同屬AAIF堆疊裡的不同層。

為什麼這對開發者是好消息？

過去多代理系統最大風險，是每家都想把自己的協定變業界標準，開發者要嘛選邊站、要嘛自己寫轉接層。A2A加入AAIF後：

治理中立化：AWS、Anthropic、Google、Microsoft、OpenAI都是白金會員，同桌訂規則
降低鎖定風險：不會因某家改變策略就整套作廢
減少整合成本：不必為每家私有協定各寫轉接程式

小克實測心得

協定整合聽起來枯燥，但對建多代理系統的工程師是實質利多。我拿MCP接工具、A2A串代理有段時間了，最大困擾是「這協定半年後還有人維護嗎」——現在有Linux基金會背書、五大廠商掛名，穩定性明顯提升。但協定成熟不代表工具鏈成熟，實際串接還是常踩SDK版本不同步、文件過時的坑，該花的除錯時間一分鐘都少不了。

好不好用，試了才知道。

🇺🇸 A2A Protocol Review: Joins MCP at Linux Foundation

One of the biggest headaches in building AI agents has been protocol fragmentation — every vendor pushing its own standard. That's starting to change. Google's A2A protocol (Agent2Agent) officially joined the Linux Foundation's Agentic AI Foundation (AAIF) in August 2026, landing in the same neutral governance body as Anthropic's MCP (Model Context Protocol). It's a real step toward agent interoperability becoming standardized instead of vendor-locked.

What Is the A2A Protocol?

A2A is an open standard that lets AI agents built by different vendors discover and talk to each other. Using JSON-LD "Agent Cards," an agent advertises its capabilities, skills, and endpoints so other agents can find it, delegate tasks, and exchange messages — regardless of who built it or what framework it runs on.

A2A vs. MCP: What's the Difference?

Short version: MCP connects agents to tools, A2A connects agents to agents. MCP standardizes how a model safely calls external databases, APIs, and file systems. A2A standardizes how independent agents coordinate with each other on multi-step tasks. Both now sit under AAIF, officially framed as layers of one stack: AGENTS.md for instructions, goose as the runtime, MCP for tool connectivity, and A2A for agent-to-agent interoperability.

Why Does This Matter for Developers?

The old risk with multi-agent systems was betting on whichever company's protocol might become the de facto standard — and getting stuck rewriting everything if it didn't. With A2A now under AAIF, that risk looks different:

Neutral governance: AWS, Anthropic, Google, Microsoft, and OpenAI are all Platinum members sitting at the same table
Lower lock-in risk: agent systems built on A2A or MCP won't be orphaned if one vendor changes strategy
Less integration overhead: no need to write a custom adapter for every vendor's proprietary protocol

Kit's Honest Take

Protocol governance news sounds dry, but if you're actually shipping multi-agent systems, this is a real win. I've been wiring MCP for tools and A2A for agent handoffs for a while, and my biggest worry was always "will anyone still maintain this in six months?" Having the Linux Foundation as a neutral steward, with five major AI vendors formally on board, meaningfully de-risks that. One honest caveat: mature governance doesn't mean mature tooling — you'll still hit SDK version mismatches and stale docs wiring things up in practice. Budget the debugging time.

好不好用，試了才知道 — the only way to know if it's good is to actually try it.

Sources / 資料來源

常見問題 FAQ

A2A協定跟MCP是競爭關係嗎？

不是。MCP負責代理連工具，A2A負責代理連代理，兩者現在同屬Agentic AI Foundation旗下的互補協定，各自解決不同層的問題。

A2A協定是誰開發的？

A2A（Agent2Agent）最早由Google開發，2026年8月正式成為Linux基金會Agentic AI Foundation的託管專案，交由中立組織治理。

開發者現在該用A2A還是MCP？

看需求：要讓AI呼叫外部工具、資料庫、API用MCP；要讓多個獨立AI代理互相協作、委派任務則用A2A，兩者可以並用。

Agentic AI Foundation底下還有哪些專案？

目前有MCP、A2A、Block的goose執行環境、OpenAI的AGENTS.md規範，以及agentgateway，官方將其定位為同一套代理堆疊的不同層。

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士