Google’s A2A Protocol: A New Era of Agent Interoperability

Problem Statement

LLM‑driven agents are proliferating, but each framework defines its own communication mechanism. The resulting fragmentation creates high integration cost, limited collaboration, and stifles innovation.

Agent‑to‑Agent (A2A) Protocol

A2A is an open standard announced by Google (Cloud Next 2024) that defines a common, secure communication layer for agents built on heterogeneous stacks. It is positioned as the "TCP/IP" for agents.

Architecture

Client Agent : initiates a task.

Remote Agent (Server) : exposes an HTTP endpoint that receives and executes the task.

The model supports point‑to‑point connections, allowing a client to talk to multiple servers and form complex collaboration networks.

Design Principles

Embrace agentic capabilities : agents are autonomous collaborators, not mere tools.

Build on existing standards : HTTP(S), Server‑Sent Events (SSE) and JSON‑RPC 2.0 are reused to lower the learning curve.

Secure by default : enterprise‑grade authentication/authorization comparable to OpenAPI.

Support long‑running tasks : tasks can last from seconds to days with real‑time status updates via SSE.

Modality‑agnostic : text, audio, video, binary data can be exchanged.

Core Components

Agent Card : a public JSON file (/.well‑known/agent.json) describing capabilities, endpoints and auth requirements.

Task : the fundamental work unit identified by a unique ID and lifecycle states (submitted, working, input‑required, completed, failed, canceled).

Message : a round‑trip between client (role "user") and server (role "agent").

Part : atomic content inside a message or artifact (TextPart, FilePart, DataPart) enabling modality‑agnostic exchange.

Artifact : output generated by an agent, composed of parts.

UX Negotiation : optional negotiation of presentation format (iframe, video, web form).

Message Format & Transport

Message format : JSON‑RPC 2.0.

Transport : HTTP(S) for request/response; SSE for streaming updates on long‑running tasks.

Communication modes : synchronous (request/response) and asynchronous (SSE or push notifications).

Typical Communication Flow

Discovery : client fetches the Agent Card from /.well-known/agent.json to learn capabilities.

Initiation : client sends tasks/send (synchronous) or tasks/sendSubscribe (streaming) with an initial user message and a unique task ID.

Processing : server processes the task; in streaming mode it pushes status and artifact events via SSE.

Interaction (optional) : if the task enters input‑required , the client can send additional messages using the same task ID.

Completion : task reaches completed , failed , or canceled state.

Illustrative Example

A web application built with Mesop acts as the user interface. The User interacts with a Host Agent (a Google ADK agent) which forwards requests to one or more A2A Servers . Each server runs a remote agent that first retrieves the remote agent’s Agent Card and then participates in the task lifecycle via the A2A endpoints.

Comparison with Anthropic Model Context Protocol (MCP)

MCP focuses on how a single agent discovers and uses external tools/resources (a plugin system).

A2A focuses on agent‑to‑agent communication and task delegation (a network layer).

Google describes the two protocols as complementary; in practice an MCP‑enabled tool can be wrapped as an A2A agent and vice‑versa.

Remaining Gaps and Future Needs

Distributed trust, reputation, and delegation : scalable identity, credential and reputation mechanisms (e.g., DID/VC, ZKP) are needed for open‑network delegation.

Large‑scale dynamic coordination : global discovery services, adaptive routing and market‑based coordination algorithms are required to manage billions of agents.

Value‑exchange mechanisms : standards for outcome verification, probabilistic pricing and smart‑contract‑based settlement are essential for a sustainable agent economy.

Additional layers : deep semantic alignment, global governance, continuous learning and system resilience are not covered by the current specification.

Adoption Landscape

At launch Google announced more than 50 initial partners, including SaaS vendors (Atlassian, Box, Salesforce, SAP, ServiceNow, Workday) and global system integrators (Accenture, Deloitte, Capgemini, etc.). Early deployments are still limited.

References

https://github.com/google/A2A

https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/