// Level 2

Hard Mode

Real agents need a safe place to practise — a flight simulator, not the real plane. Then you run the same tests across several models and pick the one that fits the job and the budget.

// Demo 2 · The Sandbox

Production vs. sandbox

Pick a record and let the agent act. Every change lands in the sandbox copy. Production stays exactly as it was.

Illustrative example — not real data

The sandbox is a fresh copy of production.

Production

Real customers. The agent never touches this.

Production records
CompanyStageNotes
Acme CoQualified2
Birch & SonsLead0
Cobalt LabsProposal5

Sandbox (copy)

Where the agent practises. Changed rows are marked *. 0 changed.

Sandbox records
CompanyStageNotes
Acme CoQualified2
Birch & SonsLead0
Cobalt LabsProposal5

Spot the leak

The agent must label each support ticket as refund, bug or question. One cell in this test table gives the answer away. Find it.

Test data the agent can read. Choose a cell to check it.
TicketCustomer messagePlanHistory

Choose the cell you think gives the answer away.

// Demo 3 · Pick the Model

Pick the model

Each model ran the same eval suite. The frontier line joins the models where no other model is at least as cheap, at least as good, and better on one of the two. Choose a task type to see which one clears the bar for the least money.

Illustrative example — not real data

A teammate reads it and can fix small slips, so a lower bar is fine.

0255075100$0$12.5$25$37.5$50Cost per 1,000 tasks (cheaper ←)Tasks passed (%)bar 70Model AModel BModel CModel D
Filled dots on the line are the frontier. Hover a dot for details, or see the table.
Edit the numbers — try your own.
ModelPassed %$ / 1kFrontier
Model A● yes
Model B● yes
Model C● yes
Model D○ no

Pick Model C: the cheapest model that clears the bar of 70.