Subscribe
AI
5 min read

Nobody Won the Agent Protocol War. A Nonprofit Did.

Eighteen months ago the plumbing for AI agents looked like the setup for a standards war. Anthropic launched the Model Context Protocol (MCP) in November 2024. Google answered with Agent2Agent (A2A) in April 2025. IBM pushed its own Agent Communication Protocol. Browser vendors started their own experiments. Every major lab wanted to own the layer where agents connect to tools and to each other. That war is over, and nobody on the founding side won it outright. The consolidation ended at a nonp

Nobody Won the Agent Protocol War. A Nonprofit Did.

Eighteen months ago the plumbing for AI agents looked like the setup for a standards war. Anthropic launched the Model Context Protocol (MCP) in November 2024. Google answered with Agent2Agent (A2A) in April 2025. IBM pushed its own Agent Communication Protocol. Browser vendors started their own experiments. Every major lab wanted to own the layer where agents connect to tools and to each other.

That war is over, and nobody on the founding side won it outright. The consolidation ended at a nonprofit instead. As of late October 2026, nearly every significant agent interoperability standard sits under the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation. The browser layer is being standardized separately at the W3C. The coordination infrastructure for machine-to-machine work is now governed by the same class of institution that governs Linux, Kubernetes, and Node.js.

The open agent stack

A timeline of the consolidation

The sequence moves fast, and the direction is one-way:

  • April 9, 2025. Google announces A2A at Cloud Next with more than 50 technology partners, including Atlassian, Box, Cohere, Salesforce, SAP, ServiceNow, and Workday. The protocol defines how agents from different vendors discover each other, delegate tasks, and exchange results over HTTP and JSON-RPC.
  • June 2025. Google donates A2A to the Linux Foundation, alongside founding organizations including AWS, Cisco, Microsoft, Salesforce, SAP, and ServiceNow.
  • August 2025. IBM's Agent Communication Protocol merges into A2A. Competing specs began folding together before the governance handover was even complete.
  • December 9, 2025. The Linux Foundation announces the Agentic AI Foundation. Three founding projects arrive at once: Anthropic's MCP, Block's open-source agent goose, and OpenAI's AGENTS.md convention. Platinum members read like a conference badge: Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI. Anthropic donates MCP the same day, citing more than 10,000 active servers and 97 million monthly SDK downloads.
  • March 2026. A2A ships its v1.0 stable specification, adding signed agent cards, multi-protocol bindings, version negotiation, and multi-tenancy.
  • April 9, 2026. At its one-year mark the Linux Foundation reports A2A backed by more than 150 organizations, with integrations across Google Cloud, Azure AI Foundry, and Amazon Bedrock AgentCore and active production deployments.
  • August 17, 2026. A2A joins the AAIF, in the same fund as MCP.
  • October 22, 2026. The AAIF adds four more projects: Anthropic's Agent Skills, OpenAI's Open Responses, Hugging Face's OpenEnv, and Alibaba's OpenSandbox. They join A2A, MCP, goose, AGENTS.md, Agent Router, and agentgateway. "A year ago, AAIF launched with three projects and a bet that agentic AI needed a neutral home to thrive. Today we host 10," said Manik Surtani, the foundation's CTO and co-founder.

The one layer outside the AAIF is the browser. WebMCP, a proposed standard that lets pages expose typed, callable tools to in-browser agents through a navigator.modelContext API, was published as a W3C Community Group draft in February 2026, edited by engineers at Google and Microsoft. Chrome ran an early preview in February and opened a public origin trial in Chrome 149 after Google I/O 2026. At TPAC in October 2026, W3C members reached broad consensus on chartering a dedicated Agentic Web Working Group. Different venue, same principle: no single vendor keeps the keys.

Why governance settled the war, not technical merit

The specs overlapped, but the deciding move was administrative. A2A's own AAIF announcement put it plainly: the protocol's 150-plus partner organizations "include direct competitors. That breadth only holds together under governance that no single participant controls."

Three forces made neutral governance the only stable end state.

Procurement reality. Enterprises will not bet their integration layer on one vendor's roadmap. Azure, Bedrock, and Google Cloud all shipped A2A support. That is only credible once Google stopped owning the spec. The same logic drove Anthropic's MCP donation: the protocol was already cross-vendor in practice, and single-company stewardship became the main objection to treating it as permanent infrastructure.

Consolidation dynamics. IBM merging ACP into A2A in August 2025 showed that late entrants could not win the interoperability layer alone. Once the two largest implementer groups sat on the same side of one spec, a protocol war had nothing left to fight over.

The playbook is borrowed, deliberately. Linux, Kubernetes, and Node.js all taught the industry the same lesson: a foundation converts competitors into contributors and turns a company strategy into public infrastructure. The AAIF is that playbook applied to agents, running on the Linux Foundation's existing legal, trademark, and governance machinery.

The stack as it now stands

The consensus framing, shared by Google, Microsoft, and Salesforce, is that MCP and A2A are complementary rather than competing. In practice the stack looks like this:

  • Agent to tool: MCP. JSON-RPC for tools, resources, and prompts.
  • Agent to agent: A2A. Signed agent cards for discovery, tasks and messages across trust boundaries.
  • Agent to browser: WebMCP at the W3C. Pages declare structured tools so browser agents act on stated APIs instead of scraping the DOM.
  • Agent to repository: AGENTS.md, the instructions convention OpenAI contributed at the AAIF founding.
  • Agent capabilities and runtime: Agent Skills for packaged competencies, OpenSandbox for isolated execution, OpenEnv for training and evaluation, agentgateway and Agent Router for traffic.

What the consolidation does not solve

Neutral governance is necessary plumbing, not a product. Three problems survive the handover.

Security and identity are still immature. Both protocols bolted on enterprise authentication in 2026, including signed agent cards and OAuth-based access layers, but the incident history of autonomous agents keeps outrunning the standards. A shared spec does not audit the agent that implements it.

The commerce layer remains contested. Agent payment and transaction vocabulary, where pricing, offers, and payment confirmation are negotiated between agents, sits outside the settled core and is where the next round of protocol competition is most likely.

And standards bodies have no enforcement power. A protocol under a nonprofit is still just a document until enough production systems speak it. A2A has roughly 18 months of real deployment behind it. MCP spec revisions in 2026 were substantial enough that implementers complain about churn. Governance ended the war; adoption is still doing the drafting.

For builders the takeaway is simple. The interoperability layer is now commodity infrastructure, and commodity infrastructure is the wrong place to differentiate. Adopt the neutral stack for connectivity and spend the budget on what the agents actually do.

One foundation, ten projects

Sources

AILLMAI Agents