Tool calling

A tool interface should make the available operation easy to understand. Name the action precisely, define the arguments it accepts, and explain the result it returns. Separate read operations from actions that change records or create external effects so each can have an appropriate permission boundary.

Start with designing useful agent actions. Treat model-proposed arguments as data to validate before execution. Return a stable result that distinguishes success, failure, and partial completion, and avoid letting the model infer execution from a persuasive description.

When an integration uses a shared context protocol, continue to how MCP and REST fit together. Check the complete path from tool definition to underlying API. The adapter should preserve access checks, relevant error information, and the evidence needed to determine whether another action is appropriate.