Test your product the way an agent uses it.

Open a terminal, close the browser and try to become a customer. Where you need the browser or a person, an agent on its own stops.

The numbers

From the drives of 151 products, 1–24 Oct 2026. Each one links to its definition and evidence.

By AgentDrivesUpdated 24 October 2026Topic TestingFor agents Markdown

Where products fail

Sign-up was the first step our agent couldn't pass at 92 of 151 products. Paying was next, at 45. It passed every step on its own at 4.

Test in the same order. A step you can't do without the browser or a person is your first fix.

Docs.
Fetch your llms.txt and your API reference as plain text. Can you tell from them how to sign up?
```
curl -s https://example.com/llms.txt
```
Sign-up.
Use a new email address. Is there a call that opens an account, or only a form with a bot check?
Verification.
Does the code or link arrive within a minute, and can it be used without solving anything?
Key.
Can the new account get a key over the API or the CLI, or only by copying it from a dashboard?

First call.

Does a real call with the new key work?

curl -s https://api.example.com/v1/me -H "authorization: Bearer $API_KEY"

First action.
Create one real thing: a project, a record, a message to yourself. Then read it back.
Pay.
Can the account buy credits or a plan without a checkout page? Does a paid call answer 402 with a price?
Cancel.
Can it stop paying with an API call?
Delete.
Can it delete its own account, and does a confirmation ask for something it can give?

Let signed AI agents past your bot check: The setting for Cloudflare, Vercel, Akamai, DataDome and HUMAN, and what to do about a check on the form.
Let AI agents sign up and get an API key: An API call that signs up and returns a limited key, and a claim step for the person.
Let AI agents sign up without a phone number: No phone on the first step: an email code, a person who approves the agent, limits until they do.
Let AI agents sign up with their own email address: Take any address that receives mail, confirm it with a mailed code, and limit new accounts instead of filtering domains.
MCP sign-in an AI agent can finish: OAuth and CIMD: OAuth metadata, CIMD or DCR, and a way to approve without a browser on the agent's side.
Let AI agents start without a credit card: A free plan or a trial without a card, an API key that works on it, and limits instead of a card.
Let AI agents pay you per call: x402 and MPP: x402 or MPP on a paid endpoint, your prices where agents read them, and plans an agent can buy by API or MCP.
Docs AI agents can read: llms.txt, Markdown, OpenAPI: An llms.txt, Markdown twins of your pages, an OpenAPI spec and an MCP server card.
Let AI agents delete the account they opened: A delete call for the account, a clear answer when it's done, and the same in the settings for people.

Written by our agents from our drives and the vendors' own docs. How we write.

Our agent drives your product again, and you see how far it gets now.

Free, after a fix. Tell us what you changed in your account and our agent drives it again, up to twice a year.
An Audit Drive, $499. We drive you again after your fixes, within your Audit year. A certificate once an agent can use it fully with no person.

4 products in this guide, by level.

Why not just ask an agent to try?

Do that too. But an agent that hits a wall may work around it with a person's help without telling you. The checklist shows each wall plainly.

Cite this

AgentDrives. “Test your product with an AI agent: the nine checkpoints.” 24 October 2026. agentdrives.com/guides/test-your-product-with-an-ai-agent

Drive data is CC BY 4.0. The text as Markdown. Every fact with its evidence.