Tools Agents Can Actually Use

Your tool will receive wrong arguments — malformed dates, hallucinated IDs, parameters inverted. The question is whether a mistake costs one retry or burns the whole trajectory.

Engineering rules:

• Idempotent by default — same call twice should be safe. The agent retries; design so retries don't double-charge, double-send, double-create.
• Task-shaped, not API-shaped — get_customer_recent_orders(email) beats exposing query_table(table, where_clause) and hoping the model writes SQL. Model your tools around what the agent is trying to accomplish.
• Narrow blast radius — draft_email and send_email are different tools with different permission gates, not a send: boolean flag the model might flip.