A live reading of the field under the cursor. On touch, the centre of the plate.
The field at rest. Values recorded at build.
Evidence before intervention.
26 years building websites, seven of them leading engineering for media brands. I measure how a system behaves before coding agents go in, and again after. In between, they work on narrow jobs behind gates they cannot edit.
Without a baseline, nobody can tell what the agents broke from what was already broken.
Three things I find when a team starts working with agents, usually in this order. If none of them sound like your team, I am probably not the person to call.
The team adopts agents this quarter, and nobody wrote down how the system behaved before. The first incident gets blamed on someone, and nobody can prove it.
Agents open more changes than anyone can read. Without tests and rules strong enough to trust, review becomes a formality or a bottleneck.
The bill comes per seat or per token, not per finished task. So nobody can tell whether the agents cost less than the work they replaced.
Per new article on my own sites, in agent runs at API list prices: research, writing, fact-check, QA and image, not counting my review. 144 articles, to . One of them: the naked-eye Andromeda limit.
An agent that can loosen a test can make any change pass.
Before agents touch a codebase, I measure it: are the tests deterministic, are the architecture rules tested, do the CI gates catch what agents get wrong. Then the same measurements after.
Which jobs go to agents and which stay with people. A written brief per agent role, its permissions, and gates it cannot edit. I run my own sites this way, every day.
Counting is work for code: exact, repeatable, nearly free. A model takes the judgments rules cannot make, such as whether two pages answer the same query, and answers with a probability, so the doubtful cases go to a person.
Where my evidence comes from: developer, technical project manager, then seven years leading engineering for two brand verticals at a European media group, with developers in France, Serbia, Brazil and Germany.
The work there: content systems multiplied by acquisitions, search at editorial scale, consent and the audience a sales team can still sell.
Two of the sites I led engineering for. In France, Médiamétrie’s count for . In Brazil, Google Analytics over the 90 days to .
If your situation is one of these, a conversation wastes your time and mine.
Nothing publishes because an agent said it was fine.
Five sites I own and run with agents. Anything I recommend, I have tried on them first.
explorecordoba
Visits a day, highest 725
From AI assistants a day, highest 20
| Day | Visits | From AI assistants |
|---|---|---|
| 465 | 5 | |
| 512 | 12 | |
| 432 | 17 | |
| 532 | 4 | |
| 534 | 10 | |
| 544 | 10 | |
| 502 | 10 | |
| 546 | 17 | |
| 489 | 9 | |
| 584 | 10 | |
| 575 | 16 | |
| 506 | 16 | |
| 687 | 20 | |
| 523 | 6 | |
| 494 | 12 | |
| 434 | 12 | |
| 424 | 8 | |
| 649 | 14 | |
| 725 | 9 | |
| 707 | 9 | |
| 550 | 8 | |
| 516 | 5 | |
| 457 | 5 | |
| 451 | 12 | |
| 516 | 8 | |
| 470 | 11 | |
| 550 | 8 | |
| 574 | 10 |
exploregranada
darkskiesatlas
cadizexperience
sottoscocca
They are run through one content factory. Each agent has a role and a written brief. A piece of content passes validators before it publishes, every commit and push runs through gates, and a build-determinism check runs before anything deploys.
Content publishes when the gates pass, and the gates have their own tests.
Visits come from Google Analytics, which counts only visitors who accept its cookies, so it undercounts. Search figures come from Search Console. Visits from AI assistants count only those that say where they came from. Google counts its own AI answers inside search and does not separate them.
Each reading is a measurement, then the change, then the same measurement again.
They come from sites I own, so the properties are named and the numbers can be checked.
of search impressions turned into clicks
Before The day trips guide converted 1.38% of its search impressions into clicks from to , at position 6.1. Its title named no destination, and the page said six trips where it listed seven.
Change On I rewrote its search title and description to lead with the real count and name the destinations.
After 1.92% from to , at position 6.9. Up 39% against 3% for twelve comparable pages left alone.
milliseconds for Google’s crawler to get a response
Before Google’s crawler waited 164 ms on average for each response from to , and made 951 requests a day. A page missing from Cloudflare’s cache had to come from the origin first.
Change On I started warming the cache every hour from the sitemap, all 777 pages, and the shared assets once a day.
After 109 ms from to , also twelve days. Crawling rose to 1,650 requests a day in that window, though Google crawls for its own reasons too.
of fact-check records that said pass without the checks behind it
Before Of the 315 fact-check records on , 201 said “pass” and nothing else. The gate checked only that word, and an agent could write it without the fact-checker ever running.
Change On I made the gate demand the work itself: a verdict for every claim checked, and a count that matches them.
After All 751 records written since carry their verdicts. In the 30 days to , the gate stopped 156 of 656 runs before anything published.
milliseconds to paint the largest element on mobile
Before The article index took 4,015 ms to paint its largest element on , on a throttled mobile profile in the lab.
Change On I reworked its fonts and layout, then added serif fallbacks and preloaded the lead image.
After 2,555 ms on , the middle of three runs. Still 55 ms over the 2,500 ms target.
A team that ships more code and fails more deploys has not become faster.
Asked by people doing the work, answered the way I would answer them in a room. Each one is dated and attributed by role.
An engagement can take one of three shapes: a fixed-scope look at something specific, a standing advisory arrangement, or embedded with your team, building. Which one fits, how long it runs and what it costs are decided per case, after the conversation.
If you are hiring for this work, write as well.
I build every day, with agents, on the sites above. What I will not do is let agents into a system where nobody can check what they did. If there is no baseline, I build it first, and the gates come before the agents. If you want the agents without the measurements, I am the wrong person, and I will say so in the first conversation.
By day I manage engineering at a multi-site European media group. I advise in a personal capacity: my employer is not party to any engagement, and I do not take work that competes with it.
English, French and Spanish. Anything I write for you is in English unless you ask otherwise.
Write to me with the team, the system, and what you want agents to do in it. I reply within two working days. If I cannot help, I will say so, and I will try to tell you who can.
Or book a call directly.
Olivier Depiesse Agentic engineering, Córdoba, Spain LinkedIn X GitHub