Agent2Agent, or A2A, is an open protocol that lets AI agents built by different companies and frameworks find each other, describe what they can do, and hand off work between them directly — without every vendor writing a custom, one-off connection to every other vendor’s agents. Google introduced A2A in April 2025 and transferred it to the Linux Foundation two months later, so no single company controls the standard today. It works alongside the Model Context Protocol (MCP) rather than replacing it: MCP connects one agent to its own tools and data, while A2A connects one agent to another.
The problem it solves
Modern AI agents — software built to plan a multi-step task and carry it out, not just answer a question — increasingly come from different companies and different frameworks: one team’s customer-service agent, another’s logistics agent, a third’s research agent. Getting two such agents to cooperate used to mean writing a bespoke integration for that one pairing, then doing it again for the next one — the same fragmentation problem email solved with a shared mail protocol, or the web solved with HTTP. A2A gives agent builders a common language for two things: discovering what another agent can do, and handing it a task to complete, so an agent built on one company’s stack can delegate work to an agent built on someone else’s without either side needing to know the other’s internal code. A retailer’s ordering agent, for instance, can hand a shipment query to a logistics partner’s agent without either company exposing its internal systems to the other.
How it works
A2A defines three building blocks. An Agent Card is a small metadata file an agent publishes describing its capabilities and how to reach it — roughly a business card one agent hands another before they start working together. A Task is the unit of work exchanged between agents, with a defined lifecycle (submitted, working, completed, or failed) both sides can track. And a transport layer carries it all over ordinary web infrastructure — HTTP, Server-Sent Events for streaming progress updates, and JSON-RPC for the message format — secured with standard web authentication such as TLS and OAuth, rather than a proprietary connection scheme.
How A2A differs from MCP
The two protocols get mentioned together often enough to blur, but they answer different questions. MCP, introduced by Anthropic in late 2024, standardizes how a single agent reaches out to external tools, files, and data sources — it’s the agent’s arms, letting it query a database, run a calculator, or read a document. A2A standardizes how one independent agent talks to another — it’s the agent’s phone line to a colleague it doesn’t work for. A single real-world system typically needs both: a travel-booking agent might use MCP to check a company’s internal calendar, then use A2A to hand the actual reservation off to a specialized booking agent built by an airline or a travel platform. For a broader look at what these systems can do on their own, see our explainer on what AI agents actually are.
Why it matters
Without a shared handoff standard, enterprises adopting AI agents from several vendors end up with agents that can’t talk to each other — a walled-garden problem that either forces everyone onto one vendor’s ecosystem or requires custom glue code for every pair of agents that needs to cooperate. A2A is meant to prevent that: an agent built with one company’s framework can still discover and delegate to an agent built with a competitor’s, as long as both speak A2A. That’s why the standard picked up unusually broad, fast backing: within its first year, more than 150 organizations, including Microsoft, AWS, Cisco, Salesforce, and SAP, had signed on, and it moved from a Google announcement to production use across multiple major cloud platforms — a sign that agent interoperability is becoming infrastructure rather than a research demo.
How to try it
A2A is free and open source under the Apache 2.0 license, with software development kits in Python, JavaScript, Java, Go, C#, and Rust. The official A2A documentation walks through the core concepts, offers a hands-on Python quickstart, and publishes the full protocol specification for anyone building an agent that needs to publish an Agent Card — or call one.