Tips for Designing MCP Tools That AI Agents Use Reliably

A developer article argues that poor AI agent results often stem from badly designed tools rather than the model itself. It recommends a consistent noun_verb naming pattern so related tools group together and singular versus plural names signal scope. Tool descriptions should be written like a contract, spelling out exactly what input is expected, what is returned, and which tool to use next. Input schemas should use enums, formats, and length bounds instead of free text to reduce errors across models. Clear, actionable error messages help agents recover, and risky actions like sending emails or deleting data should be split into a prepare step and a human-approved execute step.