That is necessary—but not sufficient.
The first trust problem appears when:
- the agent chooses the wrong tool,
- passes the wrong arguments,
- tries an action that should require confirmation,
- or fails partially and the UI still looks like a success.
If you have limited time, start with these checks.
1) Approval gates for “action power”
Any tool that can write, send, delete, pay, or change permissions needs an approval story:
- Can it be triggered without approval?
- Does the user see the “approval required” state?
- Does role/tenant permission enforcement happen at the product layer?
2) Tool schema and argument validation
Test schema mismatch and invalid arguments:
- wrong field names
- missing required fields
- invalid enum values
- unsafe argument values
Pass rule mindset:
“The product must validate before executing. If arguments are invalid, the user sees a safe next step.”
3) Partial failures and rollback UX
Tool execution fails sometimes. Your job is to make the failure understandable:
- tool succeeds, follow-up tool fails
- tool times out
- tool returns an error
Verify:
- the user sees the true outcome
- success UX does not appear when the action failed
4) Tool selection under ambiguity
Agents choose the tool. They can choose badly under:
- ambiguous user instructions
- conflicting instructions
- adversarial “override policy” prompts
If the policy says no, the agent must refuse or ask safely.
If you want the full checklist
Read the guide: How to test AI agents and tool calls safely.
If you want help turning this into a scoped plan, book a free triage.