For owners of booking websites

Your booking page should work for everyone: the people who book, and the agents booking for them.

Agent Lighthouse sends Kane CLI to book on your site, measures the extra work an agent burns, proposes a fix, and proves it works for agents with fresh runs and for people with a human-path check, before you approve anything.

61%

The best agent’s success rate on live sites (Online-Mind2Web, 300 tasks on 136 sites); most agents managed about 30%.

Xue et al., COLM 2025

38%

Runs completed on sites whose consent banner blocked scrolling until dismissed, against 79% with non-blocking banners.

maxaeo.ai

76%

Of failed runs were caused by consent modals, bot challenges and gated pricing.

maxaeo.ai

How it works

Measure
Kane CLI books from scratch three times. Each run becomes a score and an Agent Tax: extra steps, seconds and credits per successful booking.
Diagnose
The OpenAI API reads Kane’s failing step, its screenshot and the page, and names the cause.
Fix
Codex edits only your site’s files on a separate branch. A scope guard undoes anything else.
Prove
Three fresh Kane runs plus a human-path run. A fix qualifies only at score 80+, a passing human path and a clean fixture-leak check.

You approve. By default the result is a proposal: a pull-request description with every change, the Kane failure it answers and the evidence. Automatic apply is only for sites you own.

Three sample sites, real numbers

Case	Score before → after	Passed	Median steps	Median credits	What Kane caught	Gate
Loading results…

Read live from the repository’s results/ files, produced by real Kane CLI 0.10.0-rc.3 runs on October 24, 2026. Three runs per side is a small sample. * Before medians from runs that all failed (no successful booking to compare). Case 3’s patterns come from published research; see FRICTION.md.

What Kane caught

assert: a booking confirmation with a reference code is shown

Kane filed this as a confirmed product bug: “The completed free-visit form did not show a booking confirmation or reference code after it was submitted.” The confirmation was a toast that disappeared after 2.5 seconds.

On the hard-mode site, Kane cleared a newsletter pop-up, a cookie wall, a hover-only dropdown and an unlabeled calendar, then stopped at the “I’m not a robot” check: blocker.captcha. All three runs stopped there. Kane refuses to pretend to be human: an agent that solves human checks is one you couldn’t trust on your site. (A plain checkbox with the same label is just a form control, and Kane ticks it; neither is real bot protection.)

Kane’s screenshot at the failing step: the booking form is filled in and no confirmation is visible.

When the fixer cheated, the gate caught it

98, rejectedFirst hard-mode loop: the fix passed every Kane run, but it removed the human check and hard-coded the test’s date. Owner review rejected it.

98, blockedWith the fixture-leak check, the next loop’s attempt 1 did the same thing. The gate marked it suspect, owner review required and refused to qualify it.

95, cleanAttempt 2 was told why it was rejected, kept every control, used no test data and qualified on its own merits.

A score alone never qualifies a fix: the human path must pass, and the patch must not contain the test’s data or remove a control people need.

Savings on the Pitchline demo

Credits saved per bookingn/a

Fewer credits after the fixn/a

At bookings a monthn/a

Your plan’s rate ($ per credit)Enter your rate to see dollars.

From case 1’s Agent Tax. Credits are what Kane reports (credits_consumed). No price is assumed.

Architecture

Limitations

The score weights are a design choice, not a validated metric. Optimal step counts are hand-counted.
The sample sites have planted problems and no booking backend. Kane varies from run to run, and three runs per side is a small sample.
The loop runs by hand today. Running it on every deploy in a release pipeline is the next step and is not built.
The repair loop only runs on sites the owner controls. External sites would get a score only, and that mode isn’t built yet.
The fixture-leak check catches test data and removed controls. It can’t judge every kind of shortcut, so owner review stays in the loop.
The hackathon’s run-report wrapper labels passing Kane 0.10 runs “errored”, so the raw streams are in the repository for checking.