The A2A protocol, designed by Google to let agents communicate with each other, joined a foundation backed by the Linux Foundation on 20 August, one that already governs Anthropic's MCP. That body now counts more than 250 members, including AWS, Anthropic, Cloudflare, Google, Microsoft and OpenAI. Direct competitors are placing their standards under shared governance.
We have explained what MCP is: a protocol that lets a model hook into external tools without bespoke integration. Here is the follow-up, and it matters more than it looks.
Two complementary protocols
You have to distinguish what each one solves, because they do not compete with each other.
MCP connects a model to tools: a database, a service, a file system. It is the power socket between the AI and the rest of the software.
A2A connects agents to each other. It lets an agent built by one company delegate a task to an agent built by another, with a shared vocabulary for introducing itself, negotiating and reporting back.
Bringing both under a single governance produces a coherent stack: how an AI talks to tools, and how AIs talk to each other.
Why this is unusual
Competitors placing their standards under a neutral authority is rare. The natural temptation is the opposite: impose your own protocol so the ecosystem depends on you.
If everyone is accepting this here, it is because one calculation wins out. Enterprise adoption of agents is held back by integration costs far more than by model capabilities. Every company fears being locked into a single vendor, and so it waits.
A neutral standard removes that obstacle. The pie grows for everyone, and no one risks missing the market by defending an isolated protocol.
A shared standard does not just bring interoperability. It lets patches spread, data-flow checks be standardised, and best practices be common. This week we documented an attack using invisible text and a flaw exploiting persistent memory. In a fragmented landscape, every integration reinvents its protections and repeats the same mistakes. A common standard does not remove flaws; it at least stops them being discovered ten times over.
What it does not solve
A protocol defines how systems exchange, not what they are allowed to do. The central problem of agents remains untouched: when one agent mandates another, who answers for the result?
This is exactly the work we described regarding agents that pay: establishing a verifiable trail of who authorised what, and within what limits. Protocols can carry that information; they do not decide its legal weight.
And a widely adopted standard creates a shared surface: a weakness in the protocol affects everyone at once. That is the usual flip side of standardisation.
What to take away
This consolidation is the kind of news you skip over and that matters more than a model release.
It signals that the sector has moved from demonstration to industrialisation. You do not standardise an experimental field: you standardise what you plan to deploy at scale and for the long term.
For a company building with agents, the practical consequence is direct and applies right now: favour tools that speak these protocols. That is the best protection against lock-in, and it is the advice we repeat in every form — do not choose a vendor, choose an architecture that lets you change vendors.