Google and the Linux Foundation want agents to talk, not tattle

What it does A2A is a protocol specification for autonomous AI agents that need to find one another and collaborate across framework and vendor boundaries. It standardizes discovery, negotiation, and task execution over JSON-RPC 2.0 on HTTP(S), while keeping each agent’s memory, tools, and proprietary logic strictly off-limits. The project lives under the Linux Foundation and was contributed by Google.

The interesting bit Instead of forcing agents to expose their internals, A2A uses public “Agent Cards” to advertise capabilities and connection info—like a business card that doesn’t include trade secrets. The protocol also positions itself as complementary to MCP, carving out a distinct layer for agent-to-agent collaboration.

Key highlights

• Supports synchronous request/response, SSE streaming, and asynchronous push notifications
• Agent Cards expose capabilities and connection details without leaking internal state
• Handles text, files, and structured JSON payloads
• Built with enterprise security, authentication, and observability in mind
• SDKs span Python, Go, JavaScript, Java, .NET, and Rust, each in its own repository

Caveats

• This repo holds the specification and docs; the actual SDKs and sample code live elsewhere
• Authorization schemes for Agent Cards and a dynamic QuerySkill() method are still being investigated
• Streaming reliability and push notifications are explicitly flagged for future work

Verdict Worth exploring if you need agents from different vendors or frameworks to collaborate without sharing internal state. Less relevant if your architecture is strictly single-agent—A2A is explicitly not trying to replace MCP.