A2A + MCP Orchestrator

This project demonstrates a scalable multi-agent architecture where agents dynamically discover and call each other, execute domain tools via MCP, and coordinate through a central orchestrator.

The system is capability-driven, not hardcoded. Any agent can collaborate with another as long as capabilities match.

A2A (Agent-to-Agent Protocol, proposed by Google)

A2A enables dynamic discovery and communication between agents.

• registry/ acts as the discovery backbone.
• Agents register on startup (/register) and deregister on shutdown (/deregister).
• Agents are discovered via /discover or /cards based on capabilities.
• No peer URLs are hardcoded; routing is metadata-driven.

Agent Cards

Each agent exposes GET /agent-card and registers it with the registry. The card contains:

• url
• query_endpoint
• capabilities
• optional metadata

The orchestrator and other agents fetch cards from /cards and route requests using capability matching.

Result: agents can call each other dynamically without tight coupling.

MCP (Model Context Protocol, introduced by Anthropic)

MCP handles domain-specific tool execution.

• MCP servers live in mcp_servers/.
• Agents connect using MCPToolClient.
• Tools are executed via:

session.call_tool(...)

Tool execution is separated from orchestration, making domains easily replaceable.

Orchestrator

The orchestrator/ receives user requests and routes them to the appropriate coordinating agent.

In this example:

• Requests go to trainer_agent.
• The trainer discovers and invokes muscle, cardio, or diet agents based on capability.

The orchestrator manages routing only, not domain logic.

Scalability

The fitness agents are examples. The same pattern applies to:

• Finance (risk, fraud, compliance)
• Healthcare (triage, diagnostics)
• DevOps (incident, deployment)

To extend:

1. Add a new agent under agents/
2. Define capabilities and expose agent-card
3. Register with the registry
4. Attach MCP tools if required

No architectural changes are needed to scale from a few agents to many.

Runtime

Startup instructions and test commands are available in RUNNING.md.