Five short steps on recorded runs. You watch one change to the agent, guess what it does, see what happened, decide who made the call, and leave with something you can use. Nothing to install. Your answers stay in this browser, kept separately for each path, and nothing is graded.
Two versions of the same scheduling agent talk to the same patients. In one, only the patient's own words count as a yes. In the other, the AI's reading of a yes counts.
Open the recorded conversations and read one both ways. The numbers stay hidden there until you choose to see them. Under each patient line is what the patient actually meant, which the agent never saw. Look for the moment something gets decided.
Before you see a single number. It's for you: the point is to notice what you didn't expect.
Which version books the right appointment for more patients? No answer yet.
Which one books the wrong thing more often? No answer yet.
Four checks, one at a time, then one rule places it. If code the clinic controls makes the call on every write, the clinic retains the decision in code: the code is the check on each write. If a model or a vendor makes the call, it's Delegated only when someone can stop a single decision before it's written. Rolling the rule back afterwards makes it revocable, not delegated. Any other no means Ceded. Ceding isn't wrong when it's chosen and written down. The failure is the one nobody chose.
Whether the patient agreed, in the version where the AI's reading counts
The version that books more patients also books more of the wrong thing. Neither choice is wrong. What matters is that someone made it on purpose, wrote it down, and knows who answers when it goes wrong. Nothing in either version says who that is.
Before you buy an agent, take these three questions to the vendor.
Most people stop here, and that's what the walkthrough is for. If you want the next level:
Independence. Self-funded editorial work by The CTO Advisor. Google Cloud is a client, and Gemma is a Google model. Kamiwaza is a client, and appears among the platform examples. These are named here and aren't clients: Anthropic (Claude Haiku), NVIDIA (the DGX hardware), Alibaba Cloud (the Qwen models), TypeSafe (Jev), ServiceNow, Salesforce, CoreWeave, AMD and Apple. No vendor reviewed or shaped these pages before publication.
How the runs were paid for. The Gemma runs used two NVIDIA DGX Sparks The CTO Advisor owns. The Claude Haiku 5.5 runs used the $200 monthly API credit included in Keith Townsend's paid Anthropic plan; Anthropic provided no special funding or project-specific credits. The managed per-token prices shown for Gemma come from third-party listings (OpenRouter and public price trackers), not from Google.
What this is. A simulation: synthetic patients written in code, one reference scheduling agent, a deterministic scorer. It shows how the parts compare here, not how a production clinic would behave. The platform examples are unbenched.