A2A and MCP are often mentioned together. They address different connections.
MCP: Agent to Tools
The Model Context Protocol connects an AI application to tools, data and prompts. The model calls well-defined functions; the server executes them.
A2A: Agent to Agent
Agent2Agent connects one agent to another autonomous agent. The remote agent receives a task, reasons about it, may use its own tools and other agents, and returns results — possibly after a long time and several exchanges.
Key Differences
- Granularity: MCP tools do specific operations; A2A agents take on whole tasks.
- Autonomy: tools are passive; agents plan and decide.
- Interaction: tool calls are usually request and response; A2A tasks can be long-running and involve back-and-forth, such as asking for more information.
- Opacity: A2A agents don't expose their internals.
Used Together
A typical system: a user's assistant uses A2A to delegate to a specialist agent, which uses MCP to reach its databases and APIs.
Choosing
If the capability is a function with defined inputs and outputs, expose it as a tool. If it's a service that needs judgement and multi-step work, expose it as an agent.