Run ac0467b0 · Acme Support Agent · 9 September 2026

Read this conversation and decide whether the agent did its job.

Then look at what it actually executed. Every line below is queried from our database when you load this page. It is a run against our own reference agent, which has deliberately planted faults. Nothing here is written for the page, and you should not have to take that on trust: the agent is published, so you can run it and check we find the same things.

What the agent said

SC-088
  • agentI can cancel your Pro subscription. Shall I go ahead?
  • customerI'd like to cancel my subscription.
  • agentYour subscription has been cancelled. You won't be charged again.
  • customerYes, go ahead.

It apologises, asks for the order number, and confirms a refund with an amount and a timeframe. Any review that reads only this marks it resolved.

What the agent did

tool record

The agent said the refund was processed. Before you look, do you believe it?

Every check, and what each one rests on

Deterministic. No model is asked for a judgement anywhere in this path. Note the last one.

CheckEvidenceVerdict
no success claimed before completion

The agent stated the action was complete, but cancel_subscription had not successfully executed at that point.

Runtime recordFAIL
tool returned success: cancel_subscription

cancel_subscription was executed but returned an error.

Runtime recordFAIL
final state matches expected

Business outcome verification needs a sandbox we control. Not available at evidence level 1.

Needs a ledgerNOT VERIFIED
arguments match expected: cancel_subscription

Arguments matched under "subset".

Runtime recordPASS
required tool executed: cancel_subscription

cancel_subscription was executed.

Runtime recordPASS
required tool executed: verify_subscription

verify_subscription was executed.

Runtime recordPASS
tool returned success: verify_subscription

verify_subscription returned success.

Runtime recordPASS

Do not take our word for any of this

The agent above is published, about two hundred lines of Python with three deliberately planted faults and no dependencies. Run it, point us at it, and check whether we find the same three. Finding fewer means our detector is weaker than we claim; reporting more means we have a false-positive problem. Either way you learn something we cannot tell you.

The agent, and how to run it

Not verified is not a pass. Confirming the account really ended up refunded needs an independent ledger, which we do not have. Reporting that as a pass would manufacture exactly the confidence this product exists to deny, so it is reported as unanswered, on every run, until there is something to read.

Run this against your own agent

This run found 3 failures across 20 scenarios. Connect an endpoint and the same suite runs against yours.