Agent2Agent Protocol (A2A)
Agent2Agent Protocol (A2A) is an open specification published by Google in April 2025 that defines how autonomous AI agents discover each other, exchange tasks, and coordinate across different frameworks and vendors.
Agent2Agent Protocol (A2A) is an HTTP-based, JSON-RPC communication standard designed so that independently developed AI agents can interoperate without sharing code, internal state, or a common runtime. Google released the initial specification in April 2025 with support from over 50 technology partners. While Model Context Protocol (MCP) standardizes how an agent connects to external tools, A2A standardizes how one agent delegates tasks to another agent and receives structured results back.
A2A defines an Agent Card—a JSON metadata document served at a well-known URL—that describes an agent's capabilities, supported input and output modalities, and authentication requirements. When an orchestrator agent needs to delegate a subtask, it fetches the target's Agent Card, negotiates a session, and sends a Task object containing instructions and context. The subordinate agent processes the task asynchronously and returns structured artifacts. The protocol supports both synchronous request-response and streaming Server-Sent Events interaction modes, and leaves memory and context management to each agent's own implementation.
The significance of A2A is that it enables heterogeneous agent networks. An orchestrator built on LangGraph can route a financial-analysis subtask to a specialized agent running on Vertex AI, which in turn calls a compliance-checking agent built with a third-party framework—all without custom glue code. This composability is essential for enterprise workflows where different teams or vendors own different agents and changing any single agent's internal implementation should not break the surrounding system.
As of mid-2026, A2A adoption spans major platforms including Google Agentspace, Salesforce Agentforce, and SAP, as well as several open-source agent frameworks. Active working groups continue to address trust hierarchies—how agents verify the identity and authority of orchestrators—and capability versioning to handle protocol evolution without breaking deployed agents.