Editorial draft available to read. Factual and language review is pending. Examples are illustrative.
Decision and method
Describe each tool by its effect, accessible data, destinations and recovery options. Separate reading, preparing and executing. Define input validation, volume limits and actions requiring confirmation. Authorized scope should be checked by the system executing the tool rather than depending solely on the agent remembering a written instruction during a complex conversation.
Illustrative example
Illustrative example: an agent prepares tasks from a meeting and shows a preview. Final creation is limited to the chosen project and approved count. The same capability cannot be used to send emails, alter permissions or act in another project outside the request, even if the conversation includes names or links associated with it.
Acceptance check
Test out-of-scope parameters, excessive volumes, incomplete responses and cancellation before execution. Check that failure does not trigger unauthorized compensating actions. Accept when each effect can be attributed to a valid request, limits are technically enforced and a tested mechanism can stop pending work without falsely claiming that already completed external effects were reversed.
Key takeaways
- An agent needs tools with explicit scope, validation and enforceable operational limits.