AI/TLDR

A2A

Open Agent2Agent protocol for cross-framework agent discovery and collaboration

Overview

A2A — Agent2Agent — is an open protocol that lets AI agents built on different frameworks and platforms discover each other and work together on tasks without exposing their internal memory, prompts or tools. It is one piece of the emerging agent-interop stack alongside MCP for tool/data access; A2A is the layer where agents talk to other agents.

Communication runs as JSON-RPC 2.0 over HTTP(S). Each agent publishes an Agent Card describing what it can do and how to reach it, and peers can issue synchronous requests, receive streaming updates over Server-Sent Events, or get asynchronous push notifications. Messages can carry plain text, files and structured JSON so agents can exchange real artefacts rather than just chat.

The protocol was contributed by Google to the Linux Foundation as a vendor-neutral open standard. The main repository hosts the specification and samples; official SDKs are available in Python, JavaScript, Java, .NET, Go and Rust so teams can implement A2A in their existing stack without rebuilding the messaging layer themselves.

What it does

  • JSON-RPC 2.0 over HTTP(S) — a small, well-understood wire format
  • Agent Cards advertise capabilities, endpoints, auth and supported modalities
  • Sync requests, streaming via SSE, and async push notifications
  • Carries text, files and structured JSON between agents
  • Opacity-preserving: agents collaborate without exposing internal prompts or tools
  • Official SDKs for Python, JavaScript, Java, .NET, Go and Rust

Getting started

A2A is a protocol, so getting started means picking an SDK and installing it. The Python and JavaScript SDKs are the most common starting points; the spec lives at a2a-protocol.org.

Install an A2A SDK

Pick the SDK for your stack. Python and JavaScript are shown here; Java, .NET, Go and Rust SDKs are also available.

bashbash
pip install a2a-sdk
# or
npm install @a2a-js/sdk

Publish an Agent Card

Each A2A agent exposes an Agent Card describing what it can do, how to call it, and which auth and modalities it supports. The spec defines the exact fields.

Read the specification

For the wire format, message types and capability negotiation, see the canonical spec at a2a-protocol.org/latest/specification.

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

When to use it

  • Connect agents from different frameworks (LangChain, AutoGen, Semantic Kernel, …) so they can hand off tasks
  • Build hierarchical or sequential multi-agent workflows where each agent is its own service
  • Let internal agents discover and call partner agents across organisations without sharing prompts or tools
  • Add agent-to-agent interop to an existing application without inventing a custom RPC layer

How A2A compares

A2A alongside other open-source multi-agent systems tools AI/TLDR tracks, ranked by GitHub stars.

ToolStarsWhat it does
MetaGPT★ 69.5kA multi-agent framework that models a software company, assigning roles like product manager, architect, and engineer to generate code from a single prompt.
AutoGen★ 59.9kMicrosoft Research's framework for building applications where multiple agents converse with each other and with tools to solve tasks.
CrewAI★ 55.9kA framework for assembling teams ('crews') of role-playing agents that divide tasks and collaborate to complete a goal.
AgentScope★ 28.1kA framework for building multi-agent applications with message passing, visual debugging tools, and distributed execution.
OpenAI Agents SDK★ 28.1kOpenAI's lightweight Python SDK for building multi-agent workflows using explicit handoffs, tools, and guardrails.
A2A★ 24.9kOpen Agent2Agent protocol for cross-framework agent discovery and collaboration
OpenAI Swarm★ 21.8kAn educational, lightweight framework from OpenAI for experimenting with multi-agent coordination through handoffs and routines.
PentAGI★ 21.1kPentAGI is a self-hosted AI security platform that plans and runs penetration tests autonomously using a team of agents and 20+ built-in pentesting tools.