Build a Multi-Agent System With MCP — Typed, Transparent, Production-Ready

You will build a Glass-Box Context Engine with MCP as the orchestration layer. The system has an orchestrating agent that decomposes tasks and dispatches to specialised MCP servers: a retrieval agent that provides high-fidelity RAG, a synthesis agent that produces cited responses, a validation agent that checks citation coverage and domain constraints, and a moderation agent that applies safeguards before final output. Every component communicates through typed MCP interfaces.

Each specialised capability in the multi-agent system becomes its own MCP server with a coherent set of related tools. The retrieval server exposes tools for knowledge base queries and episodic memory access. The synthesis server exposes tools for response generation with different output formats. The validation server exposes tools for citation checking and domain constraint validation. This one-server-per-capability structure makes each component independently deployable and testable.

The orchestrating agent uses MCP's capability discovery mechanism to query each agent server's available tools, their schemas, and their descriptions. This discovery happens at orchestrator startup and is cached for the session. The tool descriptions are critical: they must be accurate enough for the orchestrator's planner LLM to make correct routing decisions. The workshop covers writing tool descriptions that enable reliable orchestrator routing.

Adding a new specialised agent to an existing MCP multi-agent system involves creating a new MCP server with the agent's tool definitions, registering the server address with the orchestrator's configuration, and optionally updating the orchestrator's planner prompt to describe the new capability. Because MCP provides a standard interface, the orchestrator can discover and use the new agent without code changes to the orchestration layer.

Context management across MCP agent servers uses the Glass-Box Context Engine's context routing layer: the orchestrator maintains a context store that tracks what information has been established by each agent, passes only the relevant subset of that context to each specialised agent via typed MCP parameters, and prevents context pollution by never passing one agent's full state to another agent that does not need it. This explicit context routing is what makes the multi-agent system reliable as complexity grows.

MCP adds modest network overhead compared to in-process function calls, but this is typically negligible compared to LLM inference latency. Each MCP tool invocation involves a small amount of serialization and network round-trip overhead. For high-performance systems, co-locating frequently communicating agents on the same host minimizes network latency. The workshop covers performance profiling and optimization strategies for MCP-based multi-agent systems.