A2A Protocol Unites Opaque AI Agents for Secure Collaboration
A2A uses JSON-RPC 2.0 over HTTP(S) so agents from different frameworks discover capabilities via Agent Cards, negotiate modalities like text or media, and collaborate on tasks without exposing internals, memory, or tools.
Break Silos with Opacity-Preserving Agent Communication
A2A solves interoperability for gen AI agents built on diverse frameworks like Google ADK, LangGraph, or BeeAI running on separate servers. Agents communicate as peers—not tools—via standardized JSON-RPC 2.0 over HTTP(S), enabling discovery of capabilities through 'Agent Cards' that detail skills and connection info without revealing internals. This preserves security and IP by keeping memory, logic, and tools hidden. Benefits include connecting cross-ecosystem agents for complex tasks no single agent handles alone, fostering open innovation under Apache 2.0 license via Linux Foundation with Google contribution. Trade-off: focuses on agent-to-agent, complements protocols like MCP for tool access.
Flexible Interactions for Real Workflows
Supports synchronous request/response, SSE streaming, and async push notifications for rich exchanges of text, files, JSON. Agents negotiate modalities (text, forms, media) dynamically. Enterprise design includes authentication, observability. Build workflows: expose agents as A2A servers, connect via clients, orchestrate sequential/hierarchical multi-agent systems (e.g., healthcare example across frameworks). Future adds QuerySkill() for runtime checks, dynamic UX negotiation mid-task (e.g., adding audio/video), client-initiated methods, better streaming.
Ship Fast with SDKs and Resources
Install SDKs: Python (pip install a2a-sdk), Go (go get github.com/a2aproject/a2a-go), JS (npm install @a2a-js/sdk), Java (Maven), .NET (dotnet add package A2A). Use samples repo for action. Full docs/spec at a2a-protocol.org; DeepLearning.AI course teaches compliance, connections, orchestration (taught by Holt Skinner, Ivan Nardini, Sandi Besen). Contribute via issues/discussions; 556 commits, v1.0.0 released Mar 12, 2026.