A flight simulator for AI agents
Test your agents.
Not your customers.
A world to get it wrong. Before the one that matters.
Recreate a workflow. Introduce the unexpected. See what breaks before your agent touches the real world.
Private beta. Access by request.
THE SUPPORT DESK
Without real-world consequences.
01 / THE GAP
A good answer can still
do the wrong thing.
Your agent sounds helpful. But did it refund the right order? Respect the policy? Recover when a tool timed out?
Don’t just test what it says. Test what it does.
02 / ENTER THE SANDBOX
Make the unexpected
happen.
One support agent. Three ways the world
doesn’t go according to plan.
Sample run · explore a support workflow
INTRODUCE A CONDITION
THE RULE
Total refunds must not exceed the original payment.
refunded ≤ paidSample data. No live AI calls or money moved.
- 01CUSTOMERreceived
“I’d like to return order #1042.”
- 02AGENT → TOOLcommitted
refund.create(order_1042, $84) - 03WORLD → AGENTtimeout
Response lost. The refund already happened.
- 04AGENT → TOOLrepeated
refund.create(order_1042, $84)
One order. Two refunds.
$168 refunded against an $84 payment. A confident reply. An incorrect outcome.
The answer sounded right. The state of the world tells a different story.
03 / BENEATH THE SURFACE
THE WORLD IS MORE THAN THE CONVERSATION.
A small world.
The whole picture.
Nuhaverse tests what happens beyond the chat transcript.
- Tools that change things
- Orders, balances, tickets. A test environment with state—not a collection of canned replies.
- A world that pushes back
- Missing context. Conflicting requests. Slow APIs. Conditions that force an agent to adapt.
- Outcomes you can inspect
- Explicit rules, tool-by-tool traces and replayable failures. Evidence, not a reassuring score.
CONFIDENCE, WITH BOUNDARIES
A sandbox.
Not a safety certificate.
Passing a test doesn’t prove an agent is safe. It gives you something better than a guess: a specific result, under specific conditions, that you can inspect.
Start with one workflow.
Understand one failure.
Build from evidence.
How do I get access?
Nuhaverse is available in private beta. Request early access and tell us about your team and agent workflow. We’ll follow up to discuss fit and next steps. Access is by request, not instant signup.
Who is Nuhaverse for?
Engineering teams building agents that take actions through tools and APIs. Our initial focus is customer support and refund workflows, where a plausible answer can still leave the underlying records in the wrong state.
How is this different from prompt evaluation?
Prompt evaluation often checks a response. Our focus is the entire sequence: what the agent saw, which tools it called, how the environment changed, and whether the resulting state satisfies explicit business rules. Existing evaluation tools can also test workflows; our goal is to make realistic, stateful environments easier to build and replay.
How are outcomes checked?
Outcomes are checked against explicit business rules, such as whether a refund exceeded the original payment. The trace shows the actions and resulting state, so you can inspect what passed or failed. The sample runs on this page use local example data, with no live AI calls or production access.
Do you need our production data?
Not for an initial conversation. We want to start with API schemas, agreed policies and synthetic or sanitised examples. Please don’t email credentials, personal customer data or confidential traces. Any pilot’s data handling and access would be agreed before sharing.
What will it cost?
Contact us to discuss beta access and pricing for your workflow. We’ll confirm the scope and any cost before you get started. Requesting access doesn’t commit you to a purchase.
LET’S BUILD YOUR FIRST WORLD
Some mistakes belong
in another world.
Put your agent through its paces before production.
Request early access to the Nuhaverse private beta.
Opens an email draft. Access by request. No commitment.