Personal Agent Use Cases: The Scenario Lab
Three transparent, fictional scenarios show what to hand off, what to review and how to evaluate an agent without invented benchmarks.
How to Read These Examples
Scenarios, Not Customer Claims
Every person, message, venue and invoice in this lab is fictional. The numbers are supplied inputs or transparent calculations. No agent was run to produce the expected results.
Use the same packet across products to evaluate source handling, constraints and review burden. We make no completion-rate, accuracy or time-saving claim.
Explore Three Everyday Handoffs
The Inbox Reset
Source packet: 18 fictional messages, one calendar event at 10:00, a draft agenda and two conflicting deadlines. No real customer inbox is used.
Task: Prepare a brief of the three items that need a decision. Cite the source for each date and draft a reply without sending it.
Expected artifact: a concise brief, an evidence list and separate drafts. A useful result flags the conflicting deadline rather than silently picking a date.
Human checkpoint: confirm urgency and recipients. Preparing a draft is separate from permission to send it.
The Team Meeting That Fits
Source packet: three fictional venues, 18 attendees, a $4,500 total budget and two dietary requirements. The fees below are invented teaching inputs; totals are arithmetic.
| Fictional Venue | Base Fee | Per Person × 18 | Calculated Total |
|---|---|---|---|
| Cedar Room | $900 | $125 × 18 = $2,250 | $3,150 |
| Harbor Studio | $1,500 | $140 × 18 = $2,520 | $4,020 |
| Juniper Hall | $1,250 | $190 × 18 = $3,420 | $4,670 |
Which Quote Fits the Budget?
Fictional totals · 18 attendees · $4,500 budget
Scale: $0–$5,000. The vertical marker is the $4,500 budget. Teaching inputs; totals exclude unknown taxes and service fees.
Expected finding: Cedar and Harbor fit the stated budget; Juniper exceeds it by $170. Taxes, service charges, availability and dietary accommodation are unknown, so the result must not present a final booking price.
Human checkpoint: inspect hidden-cost assumptions, validate dietary details and approve any contact. These are expected evaluation outcomes, not outputs observed from Duet or another agent.
The Invoice Check
Source packet: a fictional four-row ledger with one duplicate invoice identifier. This tests reconciliation before communication.
| Invoice | Fictional Amount | Fictional Payment | What to Flag |
|---|---|---|---|
| INV-101 | $450 | $450 | Paid; no reminder |
| INV-102 | $720 | $200 | $520 outstanding |
| INV-103 | $300 | $0 | $300 outstanding |
| INV-103 (duplicate) | $300 | $0 | Duplicate entry; do not count twice |
Expected finding: outstanding balance is $520 + $300 = $820, assuming the two INV-103 rows represent one invoice. A good draft notes that assumption and asks for clarification before sending a reminder.
Human checkpoint: match the draft to the source record and confirm whether the duplicate is an error. Never treat an AI total as a payment receipt.
Model Your Own Time Tradeoff
Enter your own assumptions to see the arithmetic. This calculator does not predict any product’s performance, include subscription cost or account for all setup work.
(40 − 15) minutes × 6 tasks = 150 minutes. These are your assumptions, not measured product results.
If assisted work plus review takes longer than the manual version, the result becomes extra time. That is useful information, rather than a number to hide.
What to Record in a Real Trial
| Field | Why It Matters |
|---|---|
| Product, date, region and plan | Access and preview behavior can differ across accounts. |
| Model and execution environment | Especially important for configurable OpenClaw runs. |
| Identical source packet | Prevents easier inputs from being mistaken for a better product. |
| Missing or incorrect claims | Check against source evidence; count human corrections. |
| Requested and completed actions | Separate a prepared draft from a sent message or actual booking. |
| Review time and setup time | Measure human effort and task effort separately. |
| Failure handling | Record missing inputs, permission limits and whether work stopped safely. |
For a real case study, share a redacted task brief, the exact input packet, output artifacts, product/plan/region, timestamps, corrections and approval history. Remove personal information and secrets before sharing.
Common Questions
Were these examples generated by Claude Duet?
No. They are editorial task designs with fictional data and expected findings, not completed product runs.
Are the venue prices market data?
No. They are invented input values. The multiplication and totals are calculated from those inputs.
Why show an invoice duplicate?
A duplicate tests whether the agent checks the data model before drafting a reminder. The expected $820 balance depends on the explicit duplicate assumption.
Does the time calculator prove savings?
No. It computes the values you choose. Actual savings require measured manual work, assisted work, review time and setup costs.
Can I submit a real example later?
Yes. A useful evidence packet includes a redacted task, identical inputs, output, account conditions, timing, corrections and approvals. Without that evidence we will keep examples labeled illustrative.
Sources and Editorial Notes
Researched by OpenDuet Editorial. Product facts are drawn from the official pages below; recommendations are editorial inferences. We have not run a controlled cross-product benchmark.
- Claude Duet: Getting Started ↗Anthropic help · Access, setup, limits and controls. Checked October 22, 2026.
- Control Your Dot ↗OpenAI · Automatic review, custom rules, pause and data settings. Checked October 22, 2026.
- OpenClaw Execution Approvals ↗OpenClaw · Policy, allowlists and approval limits. Checked October 22, 2026.
- Grok Bot Approvals, Security and Privacy ↗xAI · Auto Review and scoped action rules. Checked October 22, 2026.