Google announces the Agent2Agent protocol

Background

AI agents are programs that carry out multi-step tasks, often by calling tools and other software. Agents built by different companies, or on different frameworks, usually cannot talk to each other without custom integration work. Anthropic's Model Context Protocol (MCP) already gave agents a standard way to receive tools and context.^[1]

Consider a company that uses a scheduling agent from one vendor, a purchasing agent from another and a reporting agent built in-house. Without a shared protocol, each pair needs its own bridge, and each bridge must be updated whenever either agent changes. Shared standards let independent software work together without a bridge for each pair, and A2A follows that approach.

What happened

On 9 April 2025 Google announced Agent2Agent (A2A), an open protocol that lets AI agents built by different vendors or on different frameworks communicate, exchange information securely and coordinate actions.^[1]

Google said A2A launched with support and contributions from more than 50 technology partners, including Atlassian, Box, Cohere, Intuit, LangChain, MongoDB, PayPal, Salesforce, SAP, ServiceNow, UKG and Workday. Service providers such as Accenture, BCG, Deloitte, KPMG, McKinsey, PwC and TCS also took part.^[1]

Google released A2A as open source, with a draft specification on GitHub, and said it was working with partners on a production-ready version later in 2025.^[1]

Key details

Google described A2A as complementing MCP. MCP provides tools and context to an individual agent, while A2A handles collaboration between agents.^[1]

A2A is built on existing standards: HTTP; Server-Sent Events (SSE), which lets a server push a stream of updates over an ordinary web connection; and JSON-RPC, a lightweight format for remote procedure calls. Google said A2A offers enterprise-grade authentication and authorisation, at parity with the schemes used in OpenAPI.^[1]

The design covers everything from quick requests to long-running work lasting hours or days, with real-time status updates. It also supports modalities beyond text, such as audio and video streaming.^[1]

Long-running and real-time features matter because agent work often unfolds over time. A client that submits a task needs a way to learn about progress without repeatedly asking, which is the role of streaming status updates.

Impact

On 23 June 2025, at Open Source Summit North America in Denver, the Linux Foundation announced that it would host A2A as an open-governance project. The founding members were Google Cloud, AWS, Cisco, Microsoft, Salesforce, SAP and ServiceNow, and more than 100 technology companies supported the project.^[2]

Hosting a protocol in a neutral foundation is a common way to give a shared standard an open governance process, so that no single vendor controls it. For a developer, the practical question is which layer an integration belongs to: connecting an agent to tools or data points to MCP, while connecting one agent to another points to A2A.^[1] Whether A2A and MCP will converge or remain complementary is not settled by these announcements.