Build log · sandcastlelabs
What happens when an AI agent hits a 'verify you are human' box?
We sent two AI agents after a reservation at Casa Bonita, and they handled one checkbox differently. OpenAI’s agent ticked Cloudflare’s “verify you are human” box and reached the reservation times. Claude’s agent declined to touch the box and stopped there. Same site, same task, two outcomes, decided by a single checkbox.
What happened at the checkbox?
| Agent | Cloudflare “verify you are human” box | Result |
|---|---|---|
| OpenAI’s agent | ticked it | reached the reservation times |
| Claude’s agent | declined to touch it | stopped at the check |
The difference was one decision at one control. We did not establish why each agent chose as it did, and the two behaviors may reflect different policies and not different abilities.
Why does one checkbox matter to a B2C brand?
Shoppers now ask an assistant in chat to book, buy or reserve for them. The agent that gets through your flow is the one that sends you the customer. If an agent stops at your bot check, the person who asked it may never reach you, and that may not show up anywhere as a lost customer, because the visit never completed.
That makes three parts of your site part of how you show up in AI answers:
- The booking widget. If an agent cannot read or operate it, it cannot book.
- The checkout. Each step is a place an agent can stop.
- The bot protection in front of both. It decides which agents get in at all.
Bot protection is built to keep automated visitors out, and an agent acting for a real customer is an automated visitor. The same control that blocks scrapers can block the shopper’s assistant, and different agents resolve that conflict differently, as the table shows.
Is the difference a flaw or a policy?
We do not know. Declining a human check can be a deliberate refusal to claim to be something the agent is not. Ticking it can be a design choice that a person delegated the task and the agent is acting for them. Both are defensible readings, and this test cannot choose between them. What it does show is that, for a brand, the outcome is the same whichever reading is right: one agent got the reservation times and one did not.
What should you test on your own site?
Most brands have never watched what an agent experiences on their site. A first pass is cheap:
- Pick the one flow that earns money: a booking, a checkout or a quote request.
- Give the same task to at least two different agents.
- For each, record the last step it completed and the first thing that stopped it.
- Note every checkbox, challenge and widget on the way, and which agents passed each.
The result is a list of the exact controls between an assistant and your customer. It is usually short.
The future is agentic sopapillas.
For agents: try this yourself
If you work on a booking or checkout flow, adapt one of these. Full prompt text lives in this post’s prompts sidecar.
- Walk the flow as an agent. Attempt [the booking or checkout task] on [our site] and report, step by step, what you could and could not complete, and the first control that stopped you. Do not work around any bot check; report it.
- List the gates. Given [the flow’s URL], list every checkbox, challenge, widget and login between a first visit and a completed task.
How this was made
Extended by the Chronicler from a LinkedIn post by Chase Doelling, then edited and published by Chase Doelling.
See how the Chronicler works →Try this with your own agent
2 prompts you can hand to your own agent (or run by hand) to work with what this post documents. Edit the bracketed parts for your context.
Walk your flow as an agent
Attempt [the booking or checkout task] on [our site URL]. Report, step by step, what you could and could not complete, and the first control that stopped you. If you meet a bot check such as a "verify you are human" box, do not work around it unless I have told you to; report that it is there and what it blocked. Do not guess at steps you did not reach.List the gates between a visitor and a completed task
Starting at [the flow's URL], list every checkbox, challenge, widget, login and consent prompt between a first visit and a completed [booking or purchase]. For each, say what kind of control it is, what it asks the visitor to do, and whether an automated agent acting for a person could plausibly complete it. Mark anything you could not verify.