Google published an open protocol for agents to negotiate tasks

The conversation

On 9 April 2025, Google announced the Agent2Agent (A2A) protocol, writing that it was "launching a new, open protocol called Agent2Agent (A2A), with support and contributions from more than 50 technology partners," naming Atlassian, Box, Cohere, Intuit, LangChain, MongoDB, PayPal, Salesforce, SAP, ServiceNow and UKG among them, alongside consulting firms including Accenture, Deloitte and KPMG. Google states it "designed the A2A protocol to address the challenges we identified in deploying large-scale, multi-agent systems for our customers," and frames the new protocol as complementary to work already on file: "A2A is an open protocol that complements Anthropic's Model Context Protocol (MCP), which provides helpful tools and context to agents."

What the documents show

The protocol's own documentation site, retrieved 16 September 2026, describes the difference in mechanism: wrapping an agent as an MCP tool "is inefficient" because "agents are built to negotiate directly, and wrapping them as tools limits what they can do," so A2A instead lets an agent "publish a structured 'agent card' detailing its capabilities and contact methods," which another independently built agent can read and negotiate with directly. The reference implementation lives in a public repository released under the Apache License, Version 2.0, giving any organization a license to reuse the specification and SDKs without a separate agreement.

The system boundary

A2A's own site describes MCP as the connection between one agent and its own tools and data, and A2A as the separate layer between independent agents, each of which may itself be built on MCP internally; the launch announcement draws the same line, calling the two protocols complementary rather than competing. That boundary matters for a builder deciding which protocol governs a given integration: a call to a tool the agent itself owns is one problem, and a handoff to a different organization's agent is another.

Where it fails

A2A's own governance changed after launch in a way its site now documents: as of 16 September 2026, the protocol's blog states A2A "has officially been accepted as a Growth Stage project" at the Agentic AI Foundation, describing that body as "Linux Foundation-directed," and reports the protocol is "currently backed by over 150 organizations." That makes A2A a foundation-governed specification rather than a single vendor's proprietary format, a distinction the same post draws explicitly by naming its goal as protecting the protocol "from single-vendor constraints," though the adoption count is the foundation's and Google's own account, not an independently audited figure.

• Is a protocol's governance vendor-controlled, foundation-governed, or in transition between the two?
• Does an interoperability claim rest on a shared specification, or only on parallel announcements from cooperating vendors?
• What does an adoption figure such as "150 organizations" count, and who supplied that count?

A2A's first eighteen months moved it from a Google-authored specification with vendor partners to a foundation-hosted project, the kind of governance shift this desk treats as a fact worth naming rather than assuming, since the same words, open and standard, describe both arrangements without meaning the same thing.