Multi-Agent System: Best Practices for Context Engineering & MCP Architecture

Hi everyone,

I am working on building a multi-agent system using the Google ADK library (thanks to

's ADK video, I've been able to move forward quickly!). I am running into architectural questions and would appreciate input from the community.

1. Context Engineering Best Practices

When integrating tool calling or Model Context Protocols (MCPs), the context can become very large due to the volume of tool calling information required for complex, multi-step tasks.

What is the best practice for performing context engineering when using agents to manage large context windows filled with tool instructions and calling history?

2. MCP Architecture

Regarding MCP architecture, if the system requires many tools, what is the recommended strategy? Should we group relevant tools together and deploy multiple MCP servers, or is it better to wrap all tools within a single MCP server?

Thanks for your time and expertise on this topic.

Shima

0 comments