A2A

• A2A

  • Overview

    • Agent2Agent (A2A) is an open protocol for communication and interoperability between opaque agentic applications: agents built on different frameworks, by different vendors, on separate servers, collaborate as agents without exposing their internal state, memory, or tools.
    • Created by Google and contributed as an open source project under the Linux Foundation; Apache 2.0 licensed. In August 2026 it was accepted as a Growth Stage project of the Linux Foundation-directed Agentic AI Foundation (AAIF), alongside MCP, goose, and AGENTS.md. (announcement)

  • Normative links

    • A2A Protocol Specification — latest released spec version 1.0.0; earlier versions 0.3.0, 0.2.6, 0.1.0
    • a2aproject/A2A — spec repo; latest release v1.0.1 (2026-05-28), v1.0.0 released 2026-03-12

  • Scope

    • Agent Card — JSON metadata an A2A server publishes describing identity, capabilities, skills, service endpoint, and auth requirements; discovered at /.well-known/agent-card.json, via registries, or by direct configuration.
    • Task — the stateful unit of work with a defined lifecycle; Message — a turn with role user or agent made of Parts (text, file references, structured data); Artifact — a task output made of Parts.
    • Interaction modes: synchronous request/response, streaming over Server-Sent Events, and asynchronous push notifications to a client webhook for long-running tasks.
    • The spec is layered: a canonical data model (Protocol Buffer messages), abstract operations (send message, send streaming message, get / list / cancel task, get Agent Card), and protocol bindings for JSON-RPC 2.0, gRPC, and HTTP+JSON/REST, plus custom bindings.

  • Relationship to MCP

    • MCP connects an agent to tools and resources (vertical: deepens one agent); A2A connects agents to other agents across team or organization boundaries (horizontal). The two are designed as complements. (A2A and MCP)

  • Implementations

    • Official SDKs for Python (a2a-sdk), Go, JavaScript, Java, .NET, and Rust, linked from the spec repo README.
    • LangSmith Agent Server exposes each assistant at /a2a/{assistant_id} (methods message/send, message/stream, tasks/get) and serves its Agent Card at /.well-known/agent-card.json?assistant_id={assistant_id}. (A2A endpoint in Agent Server)
    • A LangSmith Deployment of a LangSmith/Deep Agents agent can expose it via MCP or A2A. (Going to production)