Give an LLM ten tools and it will find creative ways to call the wrong one, pass a malformed argument, or loop forever retrying a call that was never going to work.

The model isn't reading your docs, it's reading your tool names, your parameter descriptions, and your error messages, and treating all three as prompts.

Shipping MCP servers across a range of integrations, I've watched the same design mistakes tank agent reliability long before the model's reasoning is the bottleneck. This talk is about the layer most people ignore: the shape of your tools.
I'll cover concrete patterns, when to split one tool into three versus collapse three into one; writing parameter descriptions the model actually obeys; returning errors that teach the model how to recover instead of retrying blindly; designing enums and defaults so an invalid call is impossible to construct; and how to test tool ergonomics without burning a fortune on full eval runs. Everything is framework- and SDK-agnostic and applies to any MCP server in any language.