Better, or just
different?
You swap the model, rewrite the prompt, add a tool. The demo looks fine. Then a customer with two identical charges gets refunded for the wrong one.
Your agent doesn’t need another dashboard. It needs a place to live before it meets real people: the same streets, the same neighbors, every version.
Same streets, every version.
Found the town
Connect what your agent already produces. Town maps what it actually does, and you write the bylaws: what success means.
Move in the residents
Real tasks become residents. Edge cases move in on purpose: the duplicate charge, the missing order ID, the customer who changes their mind.
Let a day play out
Your agent works the town. Every action is checked against the bylaws, and every failure points to the step and the reason.
Take the census
Each new version moves into the same town. Scores line up over time, so a regression shows up before a customer finds it.
Meet the neighbors.
Built from your agent’s real traffic, not from a prompt. Every resident is a test your agent passes or fails, the same way, every version.
The refund request
Wants a refund for order 4821. Polite, has the receipt.
The plan changer
Downgrading mid-cycle, asks three times what it costs.
The double charge
Two identical charges. Agent refunded the wrong one at step 4.
The missing order ID
No order number, only an email address and a date.
Every version, counted.
Six versions of the refund agent, scored against the same 212 residents. Version 4 looked better in the demo. The census said otherwise.
Your town stays
inside your walls.
Run Town in your own cloud or in ours. Your agent’s traffic builds your town and nothing else. It is never used to train a model.
Still under construction.
Full replicas
A working copy of every system your agent touches, so you can check its actions before they reach production.
More kinds of residents
Voice calls, multi-agent handoffs and long-running tasks that span days.