Google Introduces Agent2Agent (A2A): An Open Protocol for Secure AI Agent Collaboration

Google has officially launched a new open protocol called Agent2Agent (A2A) that allows AI agents to collaborate across ecosystems securely, without being limited by any particular framework or vendor.

The protocol is supported by more than 50 technology partners—including Atlassian, Box, Cohere, Intuit, Langchain, MongoDB, PayPal, Salesforce, SAP, ServiceNow, UKG, Workday—and leading service providers such as Accenture, BCG, Capgemini, Cognizant, Deloitte, HCLTech, Infosys, KPMG, McKinsey, PwC, TCS, and Wipro.

The core goal of A2A is to enable AI agents from different sources and technologies to communicate, safely exchange information, and collaboratively execute complex tasks across enterprise platforms or applications, unlocking unprecedented efficiency and innovation potential.

A2A complements Anthropic’s Model Context Protocol (MCP); while MCP helps connect tools and resources, A2A focuses on interaction and cooperation between agents, drawing on Google’s internal experience with large‑scale agent systems.

The protocol follows five key design principles:

Embrace agentic capabilities: A2A lets agents collaborate in their natural, unstructured ways even without shared memory, tools, or context, aiming for true multi‑agent scenarios rather than simple tool usage.

Build on existing standards: It leverages widely adopted standards such as HTTP, Server‑Sent Events (SSE), and JSON‑RPC, making integration into existing IT stacks straightforward.

Secure by default: A2A supports enterprise‑grade authentication and authorization, matching the security level of OpenAPI’s auth mechanisms.

Support for long‑running tasks: The design handles both fast tasks and those that may run for hours or days, providing real‑time feedback, notifications, and status updates.

Modality agnostic: Agents can exchange not only text but also audio, video, and other media streams.

How A2A works: a “client” agent defines and transmits a task, while a “remote” agent executes it to provide information or take action. Capabilities are discovered via a JSON‑formatted “Agent Card.” Tasks have a lifecycle and can produce “artifacts.” Agents exchange messages to share context, replies, artifacts, or user commands, and each message contains “parts” with explicit content types, enabling negotiation of UI capabilities such as iframes, video, or web forms.

Example – recruiting: In a unified interface like Agentspace, a hiring manager asks a primary agent to find software‑engineer candidates based on role, location, and skills. The primary agent uses A2A to call specialized agents that search candidates, schedule interviews, and later invoke a background‑check agent, illustrating how A2A orchestrates complex workflows across multiple systems.

Reference: Google AI Blog – A2A: A New Era of Agent Interoperability