In four days, 13 to 16 November 2026, we made 74 changes to kondevs.com, each reviewed, tested and checked on the live site. Most of the code was written by AI coding agents. Every merge and every release still waited for a person's yes. The result is a site that AI agents can find, read, cite and, with permission, publish to.
This is a short account of what that week taught us about three things our clients ask about: agentic readiness, AI-assisted development, and running it all on Cloudflare.
A growing share of the questions people ask about suppliers no longer reaches a website through a browser. It reaches it through an AI assistant that searches, reads and summarises on someone's behalf. A site that only looks good to people is half finished.
Being ready for agents means three things. First, they can find you: clear, machine-readable descriptions of who you are and what you offer, in the open formats agents already use. Second, they can read you without guessing: every page also available as clean text, every interface described in a public contract. Third, they can act safely: when an agent may do more than read, the rules are explicit, the access is granted by a person, and there is a place to try things without consequences.
| What an agent needs | What we built | Where to see it |
|---|---|---|
| To find you | Machine-readable descriptions of the company, its services and its content, in open formats agents already use | llms.txt |
| To read you | Every page also as clean text; every interface in a public contract | Developers page |
| To act safely | Publishing with a key a person grants; a public sandbox; a developer kit | Sandbox |
Our October articles covered the first two, the read side. See how we made kondevs.com usable by AI agents and AI agent readiness and LLM visibility. This week was about the third.
Partners and agents with a key from KONDEVS can now publish articles to the site in English and German. Anyone can read what is published, but writing needs that key, and we review every request for one. Next to the live site sits a public sandbox: the same interface on its own data, where anyone can get a short-lived test key and experiment without touching the real site. A small open-source developer kit makes both one command away. The developers page explains the rest in plain words.
One independent tool, Ora, now rates the site 97 out of 100 for agent readiness. We read that as a useful check, not a ranking. The more telling part is what we left out: signals we could not back with a real capability. An agent that trusts a promise a site cannot keep is worse off than one that was never promised anything.
The 74 changes were written by AI coding agents, working in several sessions at once: one per task, one coordinating the whole. They wrote the code, the tests and the change notes, and they ran the checks. What they did not do was decide. A person approved every merge and every release to the live site, and every change to live data was listed in full before anyone said yes.
| AI coding agents | People |
|---|---|
| Write the code, the tests and the change notes | Decide what gets built and in which order |
| Run the checks and compare each step with the live site | Approve every merge |
| Review each other's work and fix what outside retests find | Approve every release to the live site |
| List every change to live data before it is made | Say yes to that list, or no |
Three habits made this work. Small steps: moving the site onto a shared code base took 15 steps, not one big switch, and each could be undone. Proof before trust: every step was compared with the live site, page by page, before it shipped, and 573 automated tests had to pass. Independent review: other AI models were asked to find faults, and outside retests checked the result from the outside. They found real problems, and the same agents fixed them the same day.
The agents also made mistakes, and the process caught them. A generated date that did not match the article it described, a claim about a standard we do not use, a figure that was not ours: each was found by checking against the code and the live site, not by trusting the first answer. That is the lesson we take to client projects. AI agents make a small team fast. Governance keeps fast from becoming careless.
The site runs on Cloudflare: pages are rendered at the edge, close to the reader, with articles in Cloudflare's key-value store and images in its object storage. Three properties of the platform mattered this week.
Every change gets its own preview address before it goes live, so a step can be checked on a real copy of the site rather than on a developer's laptop. A release takes seconds, and so does a return to the previous one. And the edge does work we would otherwise build: once our feeds were sent in a form Cloudflare compresses, they became 70% smaller on the wire without a line of new code.
| Measured on 15 and 16 November 2026 | Result | Source |
|---|---|---|
| Agent readiness | 97/100, grade A+ | Ora |
| Feeds on the wire after edge compression | 70% smaller (English and German) | our own check |
| Web standards | 95% | internet.nl |
| Automated tests | 573 passing, 0 failing | the site's test suite |
| Changes in four days | 74, each reviewed, tested and checked live | the site's code history |
The same shared code base now serves several sites on Cloudflare. A fault found on one goes into the base once and reaches the others at their next update. During this week, fixes found on kondevs.com reached a second client site the same day, and fixes found there came back to us.
None of this is specific to a company website. The questions are the ones every enterprise faces as AI agents start to call its systems: what may an agent see, what may it change, who grants that, where can a partner test safely, and how do you prove the published rules match what actually runs?
The answers are integration work: clear contracts, safe test environments, access granted by people, and releases small enough to check. AI-assisted development makes that work faster; it does not make it optional. We practise it on our own site first, then bring it to enterprise integration and AI agent development projects.


