A2A協議評測：Google交出治理權，跟MCP同框了 | A2A Protocol Review: Google Hands Governance to AAIF

8月 24, 2026

By Kit 小克 | AI Tool Observer | 2026-08-24

🇹🇼 A2A協議評測：Google交出治理權，跟MCP同框了

A2A協議（Agent2Agent Protocol）在2026年8月20日正式併入Linux基金會旗下的Agentic AI Foundation（AAIF），與Anthropic的MCP協議同框，由同一個中立組織治理。這代表AI代理圈裡最重要的兩套互通標準，第一次不再各自為政。這篇不吹捧，直接講對開發者實際有什麼用、有什麼還沒解決。

這次到底發生了什麼事

A2A協議原本是Google Cloud在2025年4月開源、Linux基金會在同年6月接手的專案，定位是讓不同廠商的AI代理彼此對話、分工協作的通訊格式。這次Google把A2A的治理權完整交給AAIF，跟MCP（負責代理呼叫外部工具）並列成為基金會兩大旗艦專案。AAIF從2025年12月成立至今，成員已從不到40家暴增到超過250家，白金會員包含AWS、Anthropic、Block、Bloomberg、Cloudflare、Google、Microsoft、OpenAI——幾乎所有叫得出名字的AI大廠都在同一張桌子上。

為什麼開發者該在意

過去MCP管「代理怎麼用工具」，A2A管「代理怎麼跟代理溝通」，兩套標準各自發展、各說各話，廠商之間也常常暗自角力搶主導權。現在兩者同屬一個中立基金會治理，理論上代表：

減少分裂風險：不用擔心某天Google和Anthropic的標準走向不相容的兩條路
企業採用門檻降低：Cisco、Salesforce、SAP、Oracle這類企業軟體巨頭都是Gold會員，代表企業級整合會加速
單一治理流程：規格修訂、版本控管未來走同一套治理機制，減少開發者要同時追兩套委員會動態的成本

但先別急著慶祝

老實說，基金會治理統一 ≠ 實際落地順暢。幾個現實：

OpenAI自己的AgentKit、Microsoft的Copilot Studio這類「自家生態圈」工具鏈仍在推自己的封裝層，A2A協議只是底層通訊格式，廠商鎖定的誘因並沒有消失
250多個會員名單裡有大量是「觀望型」加入，真正把A2A用在生產環境、做過跨廠商代理互通壓力測試的案例目前還很少
治理統一不等於規格穩定，A2A協議本身還在快速迭代，現在照著寫的整合明年可能要改

如果你正在做多代理系統，這次整併值得追蹤，但不用現在就把架構全部押注在A2A上——先觀察半年，看真正跨廠商互通的案例出現多少。

好不好用，試了才知道。

🇺🇸 A2A Protocol Review: Google Hands Governance to AAIF

The A2A Protocol (Agent2Agent Protocol) formally moved under the Linux Foundation's Agentic AI Foundation (AAIF) on August 20, 2026, joining Anthropic's MCP protocol under the same neutral governance. For the first time, the two most important interoperability standards in the AI agent space aren't operating as separate fiefdoms. No hype here — just what this actually means for developers, and what it doesn't fix.

What Actually Happened

A2A started life as a Google Cloud open-source project in April 2025, handed to the Linux Foundation two months later. Its job: let AI agents from different vendors talk to each other and split up work. Google has now handed full governance of A2A to AAIF, putting it alongside MCP (which governs how agents call external tools) as the foundation's two flagship projects. AAIF has grown from under 40 members at its December 2025 launch to over 250 today, with platinum members including AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI — essentially every major AI vendor at the same table.

Why Developers Should Care

MCP used to govern "how agents use tools," A2A governed "how agents talk to each other," and the two evolved separately, with vendors quietly jockeying for control of each. Consolidating both under one neutral foundation theoretically means:

Lower fragmentation risk: no more worrying that Google's and Anthropic's standards drift into incompatible directions
Lower enterprise adoption friction: Gold members like Cisco, Salesforce, SAP, and Oracle should accelerate enterprise-grade integration
Single governance track: spec revisions and versioning now run through one process, instead of developers tracking two separate committees

Don't Celebrate Just Yet

Unified governance doesn't automatically mean smooth adoption in practice. A few honest caveats:

Vendor-specific stacks like OpenAI's AgentKit and Microsoft's Copilot Studio are still pushing their own wrapper layers on top — A2A is just the underlying wire format, and the incentive to lock users in hasn't disappeared
Many of the 250+ member companies joined as observers; real production deployments that have stress-tested cross-vendor agent interop via A2A are still rare
Unified governance doesn't mean a stable spec — A2A is still iterating quickly, so integrations built today may need rework next year

If you're building multi-agent systems, this consolidation is worth tracking, but don't bet your entire architecture on A2A yet. Watch for six months and see how many genuine cross-vendor interop cases actually show up.

好不好用，試了才知道 — the only way to know if it works is to try it.

Sources / 資料來源

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