First Look at The A2A Protocol

Google, just a few days ago, announced their Agent2Agent (A2A) protocol, which aims to provide the universal handshake needed for a collaborative, multi-agent ecosystem and establish a common language for AI teamwork. Following up on last week's post about MCP (Model Context Protocol), I thought I'd take a moment to explore this new open protocol from Google and see how it compares and complements MCP.

While frameworks like LangChain, CrewAI and Semantic Kernel already provide powerful abstractions for building multi-agent systems, A2A aims to establish an open protocol for how these agents, regardless of their underlying technology, can communicate effectively. But this isn't just about enabling internal systems; the protocol's design supports discovering and interacting with agents outside an organization's own systems, fostering a truly broad and interoperable ecosystem.

A2A aims to facilitate how agents work with each other, it complements other standards like Anthropic’s MCP, which focuses on connecting agents to external tools and data sources. Put simply, MCP is for tool use, while A2A is for team work.

Google's announcment states that it developed A2A around 5 key design principles:

1. "Embrace agentic capabilities"

   The creators wanted allow agents to collaborate using their natural, unstructured modalities, enabling true multi-agent scenarios without restricting an agent to being just a "tool" for another agent.

2. "Build on existing standards" 

   By leveraging familiar technologies like HTTP, Server-Sent Events (SSE), and JSON-RPC, A2A aims to integrate more easily into the existing IT stacks businesses already rely on.

3. "Secure by default"

   Google says that enterprise-grade authentication and authorization are built-in, offering parity with standards like OpenAPI's authentication schemes.

4. "Support for long-running tasks"

   A2A was designed for flexibility, handling everything from quick requests to complex research tasks that might span hours or days, with "human in the the loop" input. Agents can provide real-time feedback, notifications, and status updates throughout.

5. "Modality agnostic"

   Recognizing that AI interaction goes beyond text, A2A supports various data types and modalities, including structured data, audio, and video streaming.

Agents advertise their capabilities via an Agent Card in standard JSON format, sort of like how a professional might advertise what they do on a business card (people still use those, right?) This allows client agents to discover potential remote collaborators. Once a suitable partner is found, they communicate using structured messages that can contain multiple parts (text, files, data, etc.), enabling rich, flexible interactions. For lengthy tasks, Server-Sent Events (SSE) allow the remote agent to stream progress updates back to the client.

A2A's open-source nature and open development alongside over 50 launch partners – including major players like Atlassian, Elastic, JetBrains, LangChain and MongoDB – signal a commitment to broad industry adoption. This collaborative approach will hopefully foster genuine interoperability, creating a level playing field where agents from different developers and platforms can effectively work together. Furthermore, the built-in discovery mechanism via Agent Cards lays the technical groundwork for a future where organizations might browse and deploy specialized agents from a vibrant "Agent Marketplace," composing sophisticated capabilities on demand.

A2A represents another step towards building more sophisticated, scalable, and truly collaborative AI systems. I look forward to playing around with it myself and testing how seamlessly it interacts and integrates with MCP servers. This was just a first look, but once I've gone a bit deeper, I might write up a more detailed post about A2A. Until then, thanks for reading.

Sources

• Announcing the Agent2Agent Protocol (A2A)