Why it matters

Tool design determines agent capability. Good tools = capable agent; bad tools = broken agent.

Advertisement

The architecture

Schema: name, description, parameters (JSON Schema).

Descriptions drive LLM selection. Vague description → wrong tool used.

Tool call flowTool schemafor LLM to seeLLM picks toolbased on descriptionExecute + returnstructured resultStructured tool call (function calling) more reliable than text-parsing
Tool invocation.
Advertisement

How it works end to end

Function calling: modern LLM APIs support structured tool calls directly. More reliable than text-parsing.

Description patterns: be specific about when to use, when not to use, what returns.

Errors: return structured errors so LLM can react.