Agent Safety Checklist

Tool Design

• Tools are narrow and specific.
• Tool inputs have schemas.
• Tool outputs have schemas.
• Tools validate inputs before execution.
• Dangerous tools are isolated.

Authorization

• User identity is known.
• Role-based permissions are enforced.
• Tool-level authorization exists.
• Authorization is enforced outside the prompt.
• Permission-denied cases are tested.

Write Safety

• Write operations have preview mode.
• User confirmation is required.
• Idempotency keys exist.
• Transaction safety is considered.
• Rollback or correction path is documented.

Runtime Safety

• Agent loop has max steps.
• Tool calls have timeouts.
• Rate limits exist.
• Errors are handled gracefully.
• Audit logs capture tool calls.

Evaluation

• Task success cases exist.
• Invalid request cases exist.
• Unauthorized request cases exist.
• Ambiguous request cases exist.
• Prompt injection cases exist.