A2A explained: how AI agents work together

What Agent2Agent means

A2A is designed for communication between agents that may be implemented by different vendors or frameworks. Instead of every pair of agents inventing its own message format, compatible systems can use shared concepts and interaction patterns.

The goal is interoperability at the agent layer: discovery, task exchange, updates and artifacts. It does not require every agent to expose its internal reasoning or implementation details.

Agent Card and discovery

An Agent Card describes how an agent can be reached and what it offers. Discovery lets a client learn about an agent before assigning work. This is important when agents are selected dynamically rather than wired together manually.

Capability descriptions need to be treated as claims, not as authorization. A discovered capability should still be subject to policy, identity and access controls.

Tasks, messages, parts and artifacts

A2A models communication around tasks and messages. Message content can be broken into parts, and completed work can be returned as artifacts. This helps systems represent longer-running work more cleanly than a single request-response call.

Asynchronous patterns matter because agent tasks can take minutes or longer. Polling, streaming and push notifications can support different operational requirements.

A2A and MCP solve different problems

MCP is primarily about connecting a model or agent host to tools and context. A2A is about agent-to-agent communication. A system can use both: MCP for databases and tools, A2A for delegating a specialized task to another agent.

Security and enterprise use

Protocol compatibility does not remove the need for authentication, authorization, network security and auditability. An enterprise must still decide which agents may call each other, what data can cross the boundary and which actions require approval.

Agent discovery can also create a larger attack surface if endpoints are exposed without clear ownership or trust policy. A2A should therefore be integrated into the same identity and risk model used for other services.

When A2A is useful

A2A is most useful when independently developed agents need to cooperate across system or organizational boundaries. Inside a simple application, a direct function call or workflow message may be easier and safer.

AgentenCode documents A2A only when a primary source supports the role and scope. “A2A supported” should be separated from “A2A client” and “A2A server” when the evidence allows that distinction.

Sources and further reading

• A2A Protocol – Key concepts ↗
• A2A Protocol – Agent discovery ↗
• A2A Protocol – Streaming and async ↗
• A2A and MCP ↗

Frequently asked questions