A2A Protocol (Agent-to-Agent Protocol)

What Is the A2A Protocol?

The Agent-to-Agent Protocol (A2A) is an open communication standard introduced by Google in April 2025 that enables AI agents built on different frameworks and by different vendors to discover each other, exchange messages, and collaborate on complex tasks. Unlike protocols that connect agents to tools and data sources—such as Anthropic's Model Context Protocol (MCP)—A2A focuses specifically on peer-to-peer agent interoperability: the ability of autonomous software agents to negotiate, delegate, and coordinate work across organizational boundaries. Backed by over 150 organizations including AWS, Microsoft, Salesforce, SAP, and ServiceNow, A2A was donated to the Linux Foundation in June 2025 and, since August 17, 2026, has been governed by the Linux Foundation's Agentic AI Foundation (AAIF). It is emerging as a foundational layer of the agentic economy.

How A2A Works: Agent Cards, Tasks, and Artifacts

A2A is built on established web standards—HTTP, JSON-RPC 2.0, and Server-Sent Events (SSE)—making it straightforward to integrate with existing enterprise infrastructure. Communication flows between a client agent that formulates requests and a remote agent that fulfills them. The protocol's core primitives include Agent Cards, JSON metadata documents that function as digital business cards advertising an agent's identity, capabilities, skills, endpoint URL, and authentication requirements. When a client agent needs to accomplish a goal, it queries available Agent Cards to identify the best-suited remote agent. Tasks are the fundamental unit of work—stateful objects with unique IDs that progress through a defined lifecycle, supporting everything from quick lookups to long-running research processes with real-time status updates. Artifacts are the outputs produced by completed tasks: documents, images, structured data, or other content composed of typed parts that can be streamed incrementally back to the client.

A2A and MCP: Complementary Layers of the Agentic Stack

A2A and MCP are not competing standards but complementary layers in the emerging agentic architecture. MCP provides the "hands"—standardizing how an individual agent accesses tools, databases, and APIs (sometimes called the southbound interface). A2A provides the "voice"—standardizing how agents communicate with each other as peers (the east-west interface). In practice, an agentic engineering stack typically uses MCP internally within each agent to access its resources, while using A2A externally to delegate work to specialized agents maintained by other teams or organizations. This division of labor mirrors the layered architecture of the internet itself, with A2A serving as something analogous to TCP/IP for autonomous software agents. Security is enterprise-grade, supporting OpenAPI-aligned authentication schemes including API keys, OAuth 2.0, and OpenID Connect Discovery.

Significance for the Agentic Economy

The A2A protocol is a critical piece of infrastructure for the agentic economy—the emerging landscape in which autonomous AI agents conduct discovery, commerce, and creative work across the open web. McKinsey projects that agent-facilitated commerce could account for $3–5 trillion in global transaction volume by 2030. Without standardized interoperability, this economy fragments into walled gardens where agents from different vendors cannot collaborate. A2A breaks down those silos by giving agents a common language for capability discovery, task negotiation, and secure information exchange. As the protocol matures—version 0.3 introduced a more stable interface for enterprise adoption—it enables increasingly sophisticated multi-agent workflows: a purchasing agent negotiating with supplier agents, a research agent delegating sub-tasks to domain-specialist agents, or generative agents in virtual worlds coordinating complex emergent behaviors across platforms. Combined with MCP, A2A is forming the nervous system of an economy where software agents are first-class economic participants.

2026: A2A Joins the Agentic AI Foundation

On August 17, 2026, A2A joined the Agentic AI Foundation, the Linux Foundation body that already hosted MCP, AGENTS.md, goose and agentgateway. AAIF says the protocol is "backed by over 150 organizations," and notes that A2A's partners include direct competitors. AAIF describes A2A as "the open standard for inter-agent communication" and "how agents discover each other, delegate tasks, and work across any framework or vendor."

The move matters less for any change to the protocol itself than for where it now sits. The two most widely adopted agent protocols—MCP, which Anthropic created, and A2A, which Google created—are now under a single neutral foundation. AAIF presents its five projects as layers of one stack: AGENTS.md for instructions and context, goose as an agent runtime, MCP for agent-to-tool connectivity, agentgateway for traffic mediation and control, and A2A for agent-to-agent interoperability. That framing settles the long-running "MCP vs. A2A" debate in favor of the complementary model described above: MCP connects an agent to its tools, A2A connects agents to each other.

The shared home also lines A2A up with MCP's evolution. MCP's 2026-07-28 specification moved to a stateless core and added a Tasks extension for long-running work, a pattern that sits naturally beside A2A's own stateful Tasks. Together they serve the long-running work agents that spread through enterprises in 2026, where one agent increasingly hands parts of a job to specialist agents owned by other teams or vendors. Commerce-specific protocols such as the Universal Commerce Protocol and AP2 remain outside AAIF, but build on the same idea of agents transacting across organizational boundaries.