TempoParking: a live parking directory built with AI agents

TempoParking is a Spanish directory of car parks, focused on free parking and on where to park near a landmark, a beach or a venue. No bookings, no payments: it lives on search. Every car park, town and landmark is a page, and a data-quality score decides which ones get indexed.

On 31 August the site had 1,836 landmark pages, 1,622 town pages and 4,840 indexable car-park pages. In twelve days of August, AI crawlers made 162,086 requests against 41,124 human page views.

Claude Code co-authored 803 of the 830 commits on the working branch: the app, the pipeline, the audits and most of the documentation. It can reach the production server, so the rules live in files it must read before acting:

• CLAUDE.md: the hard rules, including Rule 0 below.
• A decisions log, D-001 to D-081, each with an evidence level.
• One task queue: 494 items on 18 September, 317 owned by Claude, 23 waiting for me.
• A changelog: 233 dated entries with revert steps.
• 20 workflow sheets; the deploy fails if a script appears outside them.

Cursor and Codex got bounded jobs. When Codex restyled the admin panel, Claude caught its quality badges showing green for bad scores.

Every change to an agent’s rules bumps its version, stamped on each listing it writes: 3,940 listings and 5,997 revisions by 1 September. Each version states in advance how it will be judged, and the judge changed as I learned.

The quality score rewarded invention: “hours and price not confirmed” scored 93, an invented “200 m from the cathedral, open 24 h, free” scored 96. Versions are now judged by errors corrected minus true facts dropped. Since v4.0, a feature taken from reviews needs a word-for-word quote; 21 of the first 84 were paraphrases and were rejected.

A guardian blocks a deploy when an invariant breaks. Each one ships with a perturbation mode that breaks the protected thing on purpose and confirms the guardian turns red: mutation testing, aimed at the detectors.

Running them was the hard part. On 30 August, 50 of 75 guardians ran nowhere and none ran in CI. Deploys took over two hours, although only 3 of 18 failures in 30 days were real defects. Now CI runs 15 guardians on every branch, deploys gate on 58 plus 50 perturbation runs, all of them run each morning (301 on 13 September), and new failures reach the first message of each Claude session.

The project keeps about 7,400 short lessons and a search tool over them. I built a test bench to see whether the tool helped debugging agents: the same tasks with and without it, graded against written answers, over four rounds and 16 tasks with Sonnet 5 and Opus 5.

It changed zero answers and saved no more tokens than the control group; 96% of the lessons came from files the agents could already grep. Only a researched expert brief changed results, at 21–40% more tokens. Agents asked once, in two to six words, so the tool now splits each question into seven searches itself.