Frequently Asked Questions
Common Questions About Using MCP for Multi-Agent Systems
Everything you need to know before registering.
How do I decide how many MCP servers to create for a multi-agent system? +
The right number of MCP servers corresponds to the number of distinct specialised capabilities in your system, not the number of agents. Each MCP server should expose a coherent set of related tools that belong together: a retrieval server for all RAG operations, a synthesis server for document generation, a validation server for output checking. Avoid creating one MCP server per tool or one per agent interaction, as this creates unnecessary network overhead.
How do I design MCP tool schemas for effective agent communication? +
Effective MCP tool schemas for agent communication have: descriptive names that clearly indicate the tool's purpose, detailed descriptions that tell the LLM orchestrator exactly when to invoke the tool, typed parameters with validation constraints that prevent invalid invocations, and structured return types that downstream agents can parse without ambiguity. The workshop covers schema design patterns that make tools self-documenting and robust.
What is the right pattern for MCP orchestration in a multi-agent system? +
The recommended orchestration pattern uses a central orchestrating agent as an MCP client that manages the workflow, and specialised agents as MCP servers that handle specific capabilities. The orchestrator dispatches tasks by invoking MCP tools on specialised servers, collects typed results, and synthesises the final output. This hub-and-spoke pattern is simpler to debug and monitor than peer-to-peer agent communication.
How do I use MCP to pass context between agents without context pollution? +
Context pollution prevention with MCP uses two techniques: tool schema design that includes only the specific context fields each agent needs (rather than passing the full conversation state), and resource definitions that provide shared context as structured, typed objects rather than raw text. Agents request only the specific context they need through typed parameters, preventing the accumulation of irrelevant context that causes performance degradation.
How do I monitor MCP-orchestrated multi-agent systems in production? +
Monitoring MCP-orchestrated systems uses the Glass-Box logging layer to capture every tool invocation: the calling agent, the target server, the tool name, input parameters, response time, and result type. This telemetry gives you full visibility into agent coordination patterns, identifies slow tool invocations that are bottlenecking the orchestration, and tracks error rates per tool that indicate reliability issues in specific agent capabilities.
Can I use MCP with agents that run on different machines or in different containers? +
Yes. MCP is a network protocol designed for distributed deployment. Each MCP server can run in its own container or on its own machine, communicating with the orchestrating agent client over the network. The workshop covers the network configuration for distributed MCP deployments including service discovery patterns that let the orchestrator find available agent servers without hardcoding their addresses.
