Agent Tool Calling

Workflow

1. Identify agent goal, allowed tools, side effects, permissions, and stop conditions.
2. Design narrow tool schemas with clear names and typed inputs.
3. Make side-effecting tools explicit and idempotent where possible.
4. Handle tool failures, empty results, retries, and user confirmation.
5. Keep agent state and tool outputs compact and structured.

Rules

• Do not expose broad dangerous tools without guardrails.
• Return data the model can directly use.
• Separate planning, acting, observing, and final response.

Verification

• Test success, invalid input, missing permission, tool failure, and partial completion.