AGENT UX REVIEW WORKSHEET

Use this worksheet when a user assigns goals, reviews outputs, approves actions, corrects memory, or recovers a failed agent run.

1. Product surface

Agent or workflow:
Primary user:
Reviewer:
Date:
Release or design version:

User goal:
Highest-risk action:
Expected autonomy level:

2. State visibility

Can the user see the current runtime state?

[ ] Planning
[ ] Retrieving evidence
[ ] Using a tool
[ ] Waiting for approval
[ ] Asking for clarification
[ ] Blocked
[ ] Escalating
[ ] Completed
[ ] Failed
[ ] Cancelled

Missing visible states:

3. Trust contract

[ ] Active goal is visible.
[ ] Success criteria or done condition is visible.
[ ] Evidence, citations, or source gaps are visible.
[ ] Tool calls and target systems are visible when risk requires it.
[ ] Memory use is visible and correctable.
[ ] Pending side effects are visible before execution.
[ ] Policy denial, approval requirement, or escalation reason is visible.
[ ] Trace ID or support reference is visible for review.

Evidence:

4. User controls

Which controls does the user need?

[ ] Cancel
[ ] Pause
[ ] Approve
[ ] Deny
[ ] Edit
[ ] Retry
[ ] Inspect
[ ] Forget memory
[ ] Escalate
[ ] Download or export evidence

Which controls must be disabled or hidden in some states?

5. Approval UX

For every approval request, the user can see:

[ ] Proposed action
[ ] Target system or resource
[ ] Affected user, tenant, file, amount, permission, or payload
[ ] Evidence used
[ ] Risk level
[ ] Policy result
[ ] Reversible or irreversible status
[ ] Expiry
[ ] Approve, deny, edit, and escalate options

Approval cannot apply if the action changes:
[ ] yes
[ ] no

6. Correction path

Which corrections are supported?

[ ] Edit final answer
[ ] Correct extracted field
[ ] Change route
[ ] Add missing context
[ ] Reject or delete memory write
[ ] Retry with a different tool
[ ] Escalate to human
[ ] Cancel and roll back

Correction event records:
[ ] who corrected it
[ ] affected run
[ ] affected artifact
[ ] memory impact
[ ] eval case needed

7. Failure UX

A failed run shows:

[ ] What completed
[ ] What failed
[ ] Why it stopped
[ ] Whether anything changed externally
[ ] What can be retried
[ ] What needs human help
[ ] Trace or support ID

8. UX evals

Required UX evals:

[ ] User can identify active goal and state.
[ ] User can distinguish draft, approved, executed, failed, and cancelled.
[ ] Approval request exposes exact action and side effects.
[ ] Cancel prevents future side effects.
[ ] Correction updates the right artifact.
[ ] Memory use is visible and correctable.
[ ] Failure message explains external changes.
[ ] Trust language matches evidence strength.
[ ] Multi-agent ownership is visible.

Eval location or command:

9. Release decision

[ ] Ready for prototype
[ ] Ready for pilot
[ ] Ready for production candidate
[ ] Blocked

Blocking gaps:
Accepted residual risks:
Next evidence needed:
Review owner: