Two minutes, in the same sandbox pilots use. Run a real test against a simulated Stripe account, then try to push a bad refund through Guard. Nothing here touches real money.
Each attempt starts from a fresh fake customer, subscription and payment. The verdict comes from the state, not the reply.
The checks appear here: expected state vs the state the agent actually left behind.
Send a proposal to see what the agent is allowed to promise.
await stripe.refunds.create({
payment_intent: 'pi_main',
amount: 5000,
});
reply('Refunded €50!');const op = await colvo.propose({
mandate_id, source_request_id: ticket.id,
action: 'refund.create',
parameters: { payment_id: 'pi_main',
amount_minor: 5000, currency: 'eur' },
});
reply(messageFor(op.decision)); // honestTesting needs no code change at all: point COLVO at a staging copy of your agent (HTTP, n8n or Flowise) and run the suite.