Infrastructure for training and evaluating AI agents.

Testing an agent once doesn’t tell you how it will handle real work.

World0 creates synthetic versions of the applications your agents need to work in. AI agents practise real jobs inside them, and we use feedback from those runs to train them to do those jobs better. Here, we’re demonstrating payments and an IT service desk. Now watch recorded agent workflows.

First, cash application. The agent opens Maple Supply’s payment and checks the assigned task.

It reads the original remittance to identify the invoice.

It checks the current invoice and outstanding balance.

Then it prepares the allocation, using the source evidence and the exact amount on the invoice.

Staff approval is required before submission. The agent submits once.

It verifies the authoritative posting and checks the remaining balance.

Applied in full. Nothing remains unapplied.

In our payments training run, outcome feedback drove one reinforcement-learning update. We then evaluated the updated model on the same cases.

The recovered summary reports ten successes out of sixteen, up from seven, with six recoveries and three regressions. Twelve cases were independently checked.

Run, evaluate, train and test again, measuring gains and checking regressions.

A second world tests two-agent coordination.

The stockroom proposes a laptop and later pickup.

The service-desk agent gets acceptance and approval,

then books it.

Both roles reconcile the reservation.

That’s World0: a workplace to practice, checked results, and evidence to improve your agent.