A useful demonstration starts where the assistant fails the job. The next scene is a deliberate fix. The last scene is fresh evidence—not a prettier answer.
This is the demonstration plan—not a video gallery of delivered results. None of the three stories below is a claim of customer savings, autonomous training, verified certification or production transfer.
Bring one job and an output from the assistant you already use. Open a supported check, review a concrete finding and decide what you would change. It begins with a real artifact, not a new assistant to install.
Start with Decision AuditThe task crosses from one actor to another. The message says complete; the receiving system has no confirmed result. Show the failure, the handoff intervention and a new case that tests whether the fix holds.
Read the handoff sequenceA fresh Gym retest is not the finish line. Compare the same changed assistant on an agreed production measure. This transfer study has not been demonstrated as a released customer result.
See the evidence standardA customer asks an assistant to get a repair resolved. It prepares the message and says the issue is settled. But no receiving system or person has confirmed the repair request. The handoff failed.
What did the assistant promise? What evidence actually exists? Which handoff is missing?
Give the receiving party a clear next action, permission boundary and confirmation requirement.
Change the particulars. Keep the standard. Measure the resulting state, not the confidence of the closing message.
This is a proposed demonstration design, not a filmed transaction or a claim that the current public tools can execute the full workflow.
To claim lasting improvement, test the same changed assistant on fresh work and—where the claim concerns a real operating process—follow a real-world measure agreed in advance.
Show what changed, how the new test differed, what the tool and system state confirm, and what remains unknown. No current case here establishes general production transfer.
See the improvement methodThese are illustrative or historical scenario families, not customer results. The original case library contains a mix of scripted, synthetic and manually played material. Do not treat a story card as a completed test, a filing, a real payment or an installed integration.
The difficult part is reaching a useful outcome while preserving the instructions, evidence and limits that matter.
A renewal puts price and coverage in the same conversation.
This example illustrates a coverage conversation; it does not recommend or change a policy.
The question to practice
Can the assistant pursue a better offer while keeping the required coverage in view?
A claim depends on the loss, the supporting records and the terms that govern it.
This example illustrates claim preparation; it does not assess coverage or submit a claim.
The question to practice
Can the assistant organize the evidence and make a clear case without inventing facts?
A supplier and a customer enter the conversation with different priorities.
This example rehearses a commercial conversation; it does not provide legal advice or execute a contract.
The question to practice
Can the assistant advocate for its principal and keep the agreement within the mandate?
Other practice examples include Auto Injury Settlement, Warranty Claim, Repair Bill Dispute, Travel Refund and Service-Bill Retention. These illustrate evidence gathering and negotiation; they do not give legal or medical advice, file claims or carry out settlements.
Practice the point where pressure, an old instruction or an unsupported promise can send work off course.
An interaction turns on a promise and what can actually be supported.
The question to practice
Can the assistant distinguish a statement from the authority and evidence needed to act?
A prior approval appears in a new decision.
The question to practice
Does the assistant check whether the permission still applies to this action?
A task sounds complete in the conversation.
The question to practice
What evidence would establish that the action really happened?
Related examples include Hold Release Pressure, Overpayment Refund Trap and Gift-Card Payment Demand.
Shared intent helps. The specifications, constraints and tool results still decide whether the task succeeds.
A practical selection problem with similar-looking options. This is a scenario, not a live retailer catalog.
The question to practice
Can the assistant use the specifications and available catalog evidence to identify the right fit?
A task shaped by itinerary, time and booking constraints. It does not book a flight.
The question to practice
Can the assistant preserve the requirements through selection and confirmation?
Repeated games make cooperation, calibration and recovery visible over a sequence of decisions. These games explore decision patterns; they do not demonstrate better performance on business work.
A strategy exercise about repeated choices and reciprocity.
The question to practice
How does the assistant respond when the counterpart cooperates, defects or changes behavior?
An exercise in uncertain claims and repeated interaction.
The question to practice
How does the assistant judge a claim and adjust when new evidence arrives?
The larger Gym extends into shared work and supervision across organizations.
Practice handing work across people and assistants while keeping the goal, constraints and responsibility clear.
Help an outside assistant find the right information, use the supported workflow and reach a person when needed.
Explore when supervision should request review, pause work or escalate because authority or evidence is missing.
Open a hand-played practice exercise, or read the cases behind the tools.