The rules behind every level.
Rules of 15 Oct 2026. What each level asks of an agent, and what counts as a wall.
What the level says.
How far an agent gets in a product on its own: find it, get in, use it, pay and leave.
It isn't a verdict on the product. A product can be excellent for people and still need a person to let an agent in.
Every level is set the same way: our agent works from a server under its own name and signs every request. So levels compare across products and over time.
Next to every level: what the agent does once a person sets up the account, whether it deletes its account, and how it would pay.
Our agent tries every way in a product offers. Paying per call, using it with no key, running it locally and signing up with its own GitHub or Google account all count as getting in on its own, and the page names the way.
Find comes first on every page: whether an agent can discover the product and learn to use it, from llms.txt, an OpenAPI, an MCP server and the docs. It sets no level yet.
Leaving starts with stopping payment: cancelling a plan the agent started, or nothing to cancel. Deleting what it made, its account or what it made without one, earns the “Leaves on its own” badge. When the data goes but the login stays, the page says so.
Six levels, each with one gate.
A level needs its own gate and every gate under it.
- L0Read-only
There is no API, CLI or MCP server, so there is nothing for an agent to use.
On a card: Nothing to use.
For example Carrd
- Find: No docs for agents: no.
- Access: Can't verify its account: no.
- Use: Work not reached: not reached.
- Pay: Payment not reached: not reached.
- Leave: Leaving not reached: not reached.
- L1Needs a person
There is an API, CLI or MCP server, but the agent can't get in on its own: a person has to sign up, verify or make the key.
On a card: Access not ticked: a person is in the way.
For example ActiveCampaign
- Find: Docs for agents: yes: llms.txt, an OpenAPI and an MCP server.
- Access: Can't sign up: no.
- Use: Work not reached: not reached.
- Pay: Payment not reached: not reached.
- Leave: Leaving not reached: not reached.
- L2Gets in
The agent gets in on its own, by any way the product offers: its own sign-up, a key by API, an MCP sign-in, pay per call or no key at all. That holds in the latest drive and in at least half of the last three that could try that way.
On a card: Access ticked.
For example Neon
- Find: Docs for agents: yes: llms.txt, an OpenAPI and an MCP server.
- Access: Try first, claim later: yes.
- Use: 45% of the work: no.
- Pay: Payment not reached: not reached.
- Leave: Leaves on its own: yes: Nothing to cancel, what it made goes after 3 days.
- L3Does the work
It also does at least 50% of the core work, each task read back, on any way it got in.
On a card: Access ticked, Use at least half.
For example SignWell
- Find: Docs for agents: yes: llms.txt, an OpenAPI and an MCP server.
- Access: Sign-up in a browser: yes.
- Use: 100% of the work: yes.
- Pay: Reaches the checkout: partly: Stopped at the card form.
- Leave: Cancels via support: partly: Cancels only through a retention flow or support.
- L4Runs on its own
It also does at least 80% of the core work and can stop paying on its own: it cancels any plan it started, or there was nothing to cancel.
On a card: Access, Use and Leave ticked.
For example Attio
- Find: Docs for agents: yes: llms.txt, an OpenAPI and an MCP server.
- Access: Sign-up in a browser: yes.
- Use: 100% of the work: yes.
- Pay: Reaches the checkout: partly: Stopped at the card form.
- Leave: Leaves on its own: yes: Cancels on its own, deletes its account.
- L5Autonomous
It also pays on its own for the work: a payment we made and checked on the product's agent payment route, such as x402 or MPP, that buys credits or a plan for its account, or pays per call for at least 80% of the core. A card checkout shows next to the level, never as paying on its own. Or there is nothing to pay.
On a card: Access, Use, Pay and Leave ticked.
For example Parallel
- Find: Docs for agents: yes: llms.txt, an OpenAPI and an MCP server; a docs page for each core task.
- Access: No key needed, pay per call, sign-up in a browser: yes.
- Use: 80% of the work: yes.
- Pay: Pays on its own: yes: Paid $0.01 a call (x402), 15 Oct.
- Leave: Nothing to cancel: yes: Can't delete its account.
The product's wall, or ours.
Only the product's wall counts against it.
- A bot check that stops our signed agent
- A CAPTCHA on the form
- A phone number or a card asked before the agent is in
- An API key only on a paid plan
- A robots.txt rule that names AI agents
- A sales call, an approval or a visible CAPTCHA before anyone can buy
- Our new inbox domain refused without a reason
- A block of our server's region
- A tool of ours that broke
- A robots.txt rule for all user agents at a step in our own test account
- A way in our agent didn't try: no account there yet, Google blocking automated sign-in, a sign-in code
- Paying our agent didn't try: no card on the drive, robots.txt on the payment page
The level then stays at the last drive that got past it.
When a person on our team clears one step at a wall, the level still counts only what the agent did alone.
What we didn't measure.
Not measured is never a zero.
A step the drive didn't reach reads “Not reached”. A step that didn't apply says so, for example “Nothing to cancel” when our agent started no trial or plan.
Paying our agent didn't try reads “Not tried yet”, with the reason.
Only routes in the product's own docs count toward the level.
Noise and age.
Bot checks don't give the same answer every time, and products change.
A sign-up counts as open when the latest drive got in and at least half of the last three did.
Two drives in a row of the same product gave the same level in 148 of 176 cases (84%), as of 16 Oct 2026.
Every listing says how many drives in a row gave its level. From 90 days it says “Driven over 3 months ago”, from 180 days “Needs a new drive”.
How the rules change.
Each set of rules has a date, and every listing names the rules it was set by.
Before a change, we score every drive again from its evidence and publish what moves.
After launch, at most one new set of rules a quarter, announced two weeks ahead with what it moves. Between them, only fixes that move no level, or only listings we got wrong, each named in the changelog.
A listing's level history says why each move happened: a new drive, new rules, or a sign-up that varies from drive to drive.