What changed in the rules.

15 Oct 2026In force

A way in that worked before is tried again

• A lower level needs our agent to try the way in that got it in last time again, with the same account it used then.
• If our agent skipped that way, or lost it for a reason of ours, the drive doesn't count. Reasons of ours: a robots.txt rule for all user agents it still followed then, its own GitHub or Google login failing, our network.
• The second drive that confirms a lower level has to try that way too.

What it moved: 1 level moved: one listing went back from L1 to L2, to its drive from before the two drives that missed its way in.

15 Oct 2026

robots.txt inside our agent's own test account

• Our agent still follows robots.txt for every page it reads, such as docs and llms.txt.
• Inside the test account it creates, a rule for all user agents is a rule for crawlers, so our agent goes on. That covers signing up, the verification link, logging in, the key page, calls with its own key, cancelling and deleting.
• A rule that names AgentDrivesAgent, or AI agents that act for a user, still stops it, and the page shows it as the product's wall.

What it moved: No level moved on the day. Listings where such a rule stopped our agent get a new drive, and their pages change with it.

14 Oct 2026

Clearer words for paying

• Ready to pay is now Reaches the checkout: the agent got to the last step before paying and paid nothing.
• A payment our agent made that failed on the seller's side reads Tried to pay.
• Pays by card at checkout is now Card checkout.
• Nothing to pay is now No plan to buy: the product has no paid plan.
• In the API and the .md pages, the pay stage's text has the new words. Its value stays the same.

What it moved: No level moved. The words changed on 36 listings: 25 read Reaches the checkout, 6 Card checkout, 3 Tried to pay and 2 No plan to buy.

13 Oct 2026

What an agent needs: find it, get in, use it, pay, leave

• Every listing shows five outcomes: find it, get in, use it, pay and leave. Each counts whatever way the product offers: its own sign-up, a key by API, an MCP sign-in, pay per call or no key at all.
• Find comes first: 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.
• The work adds up across every way the agent got in. Without an account, the sign-up steps read Not needed.
• Paying counts for L5 when it pays for the work: credits or a plan for the account, or per call for at least 80% of the core. Less reads Pays per call for part of it.
• Leaving starts with stopping payment. Deleting what it made earns Leaves on its own, also without an account: nothing made, deleted, or gone by itself within 7 days by the docs. When the data goes and the login stays, the page says so.
• A sign-up counts as open when the latest drive got in and at least half of the recent drives that could try the same way did.
• The levels have new names: Needs a person, Gets in, Does the work, Runs on its own. The gates stay the same.
• Every listing says how an agent gets in, and the directory filters by it. Within a level, more work done comes first, then fewer minutes to the first call.

What it moved: 2 levels moved: one from L5 to L4, because it pays per call for part of the core and its account by card, and one from L1 to L2, because the drives before its latest couldn't try paying per call. 8 listings now read Leaves on its own: used without an account, they left nothing behind or what they made goes within 7 days. 3 listings read Pays per call for part of it.

13 Oct 2026

Nothing to cancel when our agent started nothing

• Cancelling is tested when our agent started a trial or a plan. When it used the free plan or paid per call, the page says Nothing to cancel.
• L4 asks the agent to cancel any trial or plan it started. A paid plan it didn't start doesn't count against it.

What it moved: 1 listing went up to L5 on a new drive; under the rules before, that drive would have been L3. The leaving result changed words on 35 listings: 17 read Leaves on its own and 18 Nothing to cancel.

12 Oct 2026

Every way in, the docs read and the checkout walked

• A drive reads the product's docs, up to 600 pages, and says what they cover: getting started, sign-in, errors, examples, how recent they are, and a page for each core task.
• A drive tries every way in the product offers, the ones made for agents first, and lists what happened on each.
• Signing up with our agent's own GitHub or Google account counts as getting in on its own, and the page names the account.
• A way in our agent didn't try is ours to try, never the product's wall.
• Paying shows one of six results: pays on its own, ready to pay, a person has to buy, not tried yet, not reached, nothing to pay. Not tried yet comes with our reason and is never shown as the product's failure; a card checkout reads as a fact.
• L5 asks for a payment we made and checked on the product's agent payment route, such as x402 or MPP. A card checkout shows next to the level, never as paying on its own.
• A drive walks the checkout up to the button that pays and stops there, and records what it saw: the checkout, a CAPTCHA, a phone field, the price.
• Each documented way to do the core work is tried once. What the docs promise and doesn't work, and what works without docs, are counted.
• Every drive the directory shows now runs up to 90 minutes.

What it moved: No level moved. The paying result changed words on 51 listings: 26 read Ready to pay, 15 Payment not reached, 7 Pays by card at checkout, 2 A person has to buy and 1 Not tried yet. Fathom reads Needs a card to start.

12 Oct 2026

robots.txt, the levels' gates and the facts at the door

• A robots.txt rule for all user agents is a rule for crawlers. When our agent stops at one on the way into its own account, the stop is ours: the level stays at the last drive that got past it, and the page says so.
• A robots.txt rule that names AI agents is the product's wall, and the page says so.
• L2 means one thing: an agent gets in on its own and does less than half the core work. What it does once a person sets up the account shows next to every level.
• L4 asks the agent to cancel any plan it started on its own. Deleting the account shows next to the level.
• L5's words promise only a payment we made and checked. A product with nothing to pay says so.
• What stopped the agent at the door is recorded from a fixed list and checked in review.

What it moved: 5 levels went up from L3 to L4. 6 listings went back to an earlier drive (2 of them to L4) and 1 waits for a new drive.

12 Oct 2026

Fixes to the rules of 11 Oct

• More routes that work without an account count as getting in.
• Where a bot check refused the sign-up form, the check is the stop, not robots.txt.
• A cancel that wasn't tried no longer gets a grade.

What it moved: 1 level went up.

11 Oct 2026

Getting in without an account, robots.txt and our new inbox domain

• A documented route that works without an account counts as getting in, and so does a paid call without an account that we made and checked.
• robots.txt that keeps agents out of a step to get in counts as the stop.
• A sign-up refused without a reason, tried only with our new inbox domain, counts as ours until a drive also tried our second domain.

What it moved: 4 levels went up. 9 listings waited for a new drive and 3 went back to an earlier drive.

8 Oct 2026

Paying on its own means a payment we made

• Paying on its own needs a payment our side made and checked. A way to pay the drive only found counts as found.

What it moved: No level moved. On 9 listings a way to pay that was only found stopped counting as paying.

8 Oct 2026

The complete drive

• When a product has paid plans and the drive started none, its cancel counts as not tested.
• A bank check on our own card is ours, never the product's.

What it moved: No level moved. 16 listings say their cancel was not tested.

7 Oct 2026

A step a person cleared

• A step a person on our team cleared at a wall counts as the person's. What the agent did after it shows in its own section, never in the level.

What it moved: No level moved.

6 Oct 2026

Getting in, in a browser

• An agent that signs up and works the product in its browser gets in, also when only the API key needs a paid plan.

What it moved: 3 levels went up.

5 Oct 2026

Our side is never the product's wall

• A drive our side stopped is never published as the product's wall.
• A step our agent did in its own browser stays its own.
• Holds up comes only from the crash test.

What it moved: 2 levels went up and 3 listings waited for a new drive.

3 Oct 2026

Names for the levels, and no overall score

• The levels got their names. The level is the only rank, with part scores next to it, never a blend.

What it moved: No level moved.

3 Oct 2026

Paid-only work

• Core work that needs a paid plan is left out of the work an agent is measured on.

What it moved: No level moved.