On August 20, 2026, the Linux Foundation's Agentic AI Foundation announced that Google's A2A protocol had joined its stable of hosted projects, sitting alongside Anthropic's Model Context Protocol, which became a founding project when the foundation launched. In under a year, the AAIF has grown from 49 founding members to more than 250, including every major cloud provider and model lab. That growth number matters less than what it now hosts. For the first time, the two protocols that define how AI agents talk to tools and how AI agents talk to each other live under the same neutral governance umbrella, instead of being controlled by the two companies that happened to invent them.
If you have shipped anything with agents in it over the last two years, you have already felt the cost of the alternative. Every team building multi-agent systems has had to make a bet on which company's protocol would still be relevant in eighteen months, and every team integrating a partner's agent has had to write a bespoke adapter because there was no shared contract for agent-to-agent communication the way there was, eventually, for agent-to-tool communication. A2A joining MCP inside a vendor-neutral foundation is the industry's clearest signal yet that both layers are now considered infrastructure, not competitive moats, and that the era of picking a protocol because of who built it rather than what it does is ending.
The two protocols are not competitors, and treating them as interchangeable is the single most common architectural mistake we see teams make when they first start building agentic systems. They solve different problems, and a production system typically needs both.
A useful mental model is that MCP is the protocol inside your agent's head, connecting it to the tools it reaches for, while A2A is the protocol between two heads, connecting your agent to somebody else's. Most production systems we have built over the last year need both: an orchestrating agent that uses MCP to pull data and call internal tools, then hands a subtask to an external or third-party agent over A2A when the work crosses a trust boundary the calling agent does not control.
Moving a protocol into a Linux Foundation project is not a cosmetic rebrand. It changes who can veto a breaking change, who funds the maintainers, and how disputes between competing implementers get resolved. Anthropic donating MCP as a founding project and Google folding A2A in a year later both give up unilateral control over the specification's roadmap in exchange for broader adoption and a governance structure that large enterprises are more comfortable building compliance programs around.
For teams actively designing multi-agent systems, this consolidation should change a few concrete decisions rather than just providing reassurance about the industry's direction.
Consolidation reduces one kind of risk while introducing a different one worth watching over the next year. A shared protocol stack used by 250-plus organizations is a much larger attack surface than any single vendor's proprietary format, and a vulnerability discovered in either A2A or MCP's core specification now has blast radius across the entire ecosystem rather than one company's product line. The MCP ecosystem already learned this lesson the hard way earlier in 2026, when a wave of critical CVEs surfaced across reference server implementations faster than most teams were prepared to patch. A2A joining the same governance umbrella means the same discipline, tracking security advisories, subscribing to the foundation's disclosure channel, and testing patches against a staging environment before they hit production agents, now needs to extend to your agent-to-agent surface as well, not just your tool-calling surface.
There is also a slower-moving risk worth naming honestly: neutral governance reduces but does not remove the influence of the companies that still employ most of the maintainers and fund most of the development hours. Watch who actually shows up to the technical steering committee meetings over the next few quarters, not just who is listed as a foundation member, before treating either protocol as fully vendor-independent.
The practical shift here is simple even if the governance story is not: A2A and MCP are no longer bets on which company wins, they are infrastructure with the kind of neutral stewardship that makes them safe to build a multi-year architecture on. If your system still routes agent-to-agent calls through a homegrown protocol because A2A felt too immature or too tied to Google when you built it, this is the moment to revisit that decision. If you have not yet drawn a clean boundary between your tool-calling layer and your agent-to-agent layer, doing that now will save a much more painful refactor once both protocols keep evolving under foundation governance.
Working out where that boundary belongs in an existing system, and sequencing a migration onto A2A without breaking agents already in production, is exactly the kind of multi-agent architecture work our AI consultancy practice does for clients. Start by mapping every place your agents currently cross a trust boundary, since that map is what decides how much of this consolidation actually applies to your stack.
Before we start, please share a few details so we can follow up with you.
End this conversation? Your chat will be emailed to us.