THE USE CASE
Where this helps.
Automate a narrow multi-step process without losing oversight. Use a small sandboxed agent to interpret inputs and draft actions that require approval.
How to try it responsibly
- 01
Specify the permitted tools, data boundaries, and success conditions.
- 02
Test happy paths, failures, and prompt-injection examples.
- 03
Add logs, human approvals, timeouts, and rollback plans.
A starting prompt
STARTING TEMPLATE / ADAPT TO YOUR CONTEXT
For this authorized request, propose actions from the approved tool list; ask for confirmation before any write.
How to verify the result
Track end-to-end success and harmful or unnecessary tool calls.
Limits and failure cases
Agents can make incorrect tool decisions; restrict access and require approvals.
Editorial note: This is a practical starting example, not an independent performance benchmark or a guarantee that a specific product can complete the workflow autonomously. Test with your own permissions, tools, and representative cases. Relevant technical/reference source ↗