A2A Joins the Agentic AI Foundation, Beside MCP

Google has handed the Agent2Agent protocol to the Agentic AI Foundation, the same outfit that already minds the Model Context Protocol. For the first time the two standards most agent builders actually touch sit under one vendor-neutral roof: MCP, which lets an agent call a tool, and A2A, which lets an agent call another agent.

A2A is barely older than a toddler. Google launched it in April 2025, donated it to the Linux Foundation, absorbed IBM’s rival Agent Communication Protocol in August 2025, and on 18 August 2026 moved it once more. The foundation reckons A2A now has backing from more than 150 organisations and runs in production across supply chains, financial services and phones. Membership of the wider foundation has gone from fewer than 40 at its December 2025 launch to more than 250. For a protocol this young, that is a lot of road under the tyres.

Table of Contents

What A2A actually does

Strip away the branding and A2A is a way for agents to introduce themselves. Each agent publishes an agent card, a structured description of what it can do and how to reach it. Another agent reads the card, works out the capability it needs, and hands over the task, no human brokering the meeting. The exchange is structured, observable and framework agnostic, so an agent written in one stack can delegate to one written in a completely different stack without a bespoke connector for that exact pairing.

Version 1.0, out in March 2026, added the enterprise-shaped bits: multitenancy, version negotiation, multi-protocol bindings and signed agent cards for cryptographic identity. That last one matters more than it sounds, and I will come back to it.

Who is already wired up

This is not a slideware standard. The list of places A2A is running in anger is genuinely broad:

Huawei standardised on it to link Celia, its OS-level assistant, with in-app agents across HarmonyOS.
Tencent’s WeChat uses it to talk to Huawei and other Android assistants.
Google Cloud, Microsoft Azure AI Foundry and AWS Bedrock AgentCore can all host or discover A2A agents.
Google Cloud and PayPal are extending it into payments through a sister spec, the Agent Payments Protocol.

Having MCP and A2A in the same foundation is the useful bit for the rest of us. Buyers can now lean on vendors to support both instead of accepting a one-supplier agent stack, and the integration tax that made multi-agent work miserable drops a notch.

The bit the press release skips

Standardising the pipe does not make the water clean. Mahesh Shanmugasundaram, lead AI solutions architect at Seekr, calls the risk an ‘AI game of telephone’. In an A2A chain, each agent tends to treat the previous agent’s output as trusted input rather than a claim to be checked. A small hallucination at step one can arrive at step five looking authoritative, because every hop added confidence and nobody added verification.

He is right, and it is structural. Interoperability widens the blast radius: the easier it is for agents to pass work around, the further a bad answer travels before anyone notices. The signed agent cards prove who an agent is. They prove nothing about whether what it said is true.

What to do with it this week

If you are building anything multi-agent, this is worth a look now rather than later. Map which of your services should publish agent cards, ask any vendor point-blank how they support A2A and MCP, and treat every cross-agent handoff as untrusted input. Evidence, confidence scores and traceability need to be the default unit of exchange between agents, not a feature you bolt on after the first embarrassing incident.

Did you know: IBM’s Agent Communication Protocol did not compete with A2A for long. It was folded into it in August 2025, a rare case of two agent standards merging instead of multiplying.

Sources

The free stack. One email a week: the AI tools and moves that actually matter, hype filtered out. Subscribe free →