What is A2A?
The Agent2Agent (A2A) Protocol is an open standard that enables AI agents to communicate and work together seamlessly.
Developed by Google, A2A creates a common language for AI agents to share capabilities and collaborate on tasks.
Its main goal is solving a critical enterprise AI challenge: allowing agents from diverse ecosystems to work together effectively.
Key Principles & Benefits
Cross-platform Collaboration
Agents can work together regardless of who built them or their underlying framework.
Simple & Standardized
Reuses existing web standards (HTTP, JSON) where possible, simplifying integration.
Enterprise-ready
Incorporates auth, security, privacy, tracing, and monitoring for business needs.
Async First
Designed for both short and very long-running tasks, including human-in-the-loop scenarios.
Modality Agnostic
Supports diverse content types like text, files, audio/video, forms, and structured data.
Opaque Execution
Agents don't need to expose their internal thoughts, plans, or tools to collaborate.
How It Works: Core Concepts
A2A uses familiar web technologies (HTTP, JSON) and defines several core concepts to facilitate agent interaction:
Agent Card
A metadata file (e.g., /.well-known/agent.json) describing an agent's capabilities, skills, endpoint, and authentication.
A2A Server & Client
An A2A Server exposes an HTTP endpoint. An A2A Client (another agent or app) consumes its services.
Task
The central unit of work, identified by a unique ID, progressing through states (submitted, working, completed, etc.).
Message
Represents communication within a task between the client (user role) and agent (agent role), containing Parts.
Part
The fundamental content unit within Messages or Artifacts (e.g., TextPart, FilePart, DataPart for JSON).
Artifact
Outputs generated by the agent during a task (e.g., files, structured data) delivered as Parts.
Typical A2A Flow
1. Discovery
Client fetches Agent Card (e.g., /.well-known/agent.json) to understand capabilities.
2. Initiation
Client sends request to Server to create a Task with the initial message.
3. Processing
Server processes the task, potentially sending streaming updates (SSE).
4. Interaction
If needed, task becomes input-required; Client sends more messages.
5. Completion
Task reaches terminal state (e.g., completed) with potential Artifacts.
6. Updates (Optional)
Server uses Push Notifications (webhooks) for updates in disconnected scenarios.
Real-World Applications
Calendars & Scheduling
Calendar agents working with scheduling and notification agents.
Document Collaboration
Document creation agents collaborating with editing agents.
Customer Service
Customer service agents connecting across different systems.
Research & Analysis
Research agents gathering information from specialized domain agents.
Complementary to MCP
While Model Context Protocol (MCP) focuses on connecting AI models with tools, A2A enables complete agents to work together as independent applications. Together, they create a comprehensive ecosystem for AI interoperability. Read more about the differences in our A2A vs MCP comparison.