What we say about the tower — click and read.
How many people must leave before nobody knows a piece of code anymore? The usual measure gets it wrong: it counts email addresses instead of people, and it mistakes large automated cleanups for real work. Once corrected, it sometimes names an owner who does not exist.
An agent on a home machine, paying nothing each month. The map first, then memory, then tools — the large model only as a last resort.
A four-legged robot is easy. A robot you want to touch is not. The skeleton is copied from a real dog, bone by bone, and the fur is fitted so that it floats — that float is what makes the little "puff" at every step.
Everyone builds agents; nobody sees them. We copied the tool everyone used at school to learn networks: the devices are the agents, the cables the pairs that talk, the packets their messages. Green when the gate accepts, red when it refuses — with the reason written beside it.
In a company, delivery speed matters more than team size. Clark, level 10, 2,525 missions delivered, is the measured proof.
One page that plays your agents' conversation like a network animation. Green means the gate said yes. Red means it refused, and it writes down why.
I stopped filling in agent forms and built a room instead. Seven parts, each one a real field. Drawing it found 271 tools attached to nothing.
Then she declares she found nothing, and wakes up her large language model.
write a test bench — a list of checks that answers yes or no with numbers: raw text goes in, a clean document comes out, verified by Lois, who proofreads.
That is the blind check we described in "A blind check that says green": it never breaks down, it reassures.
As many times as it takes, until every sentence speaks to me.
The idea is simple: instead of filling boxes, you watch the agent move along a conveyor belt and build itself, station after station.
Everything is on the table, and it gets decided as we go:
Two containers, an isolated network, nothing leaving the machine.
That is where something becomes clear without a single word of explanation: an agent missing a leg still stands up on screen, and goes nowhere.
An agent is an assembly: a memory to find its bearings, a trade, actions and refusals, tools, a brain, and a way of showing where it stands.
Seven of them went green on the first run: the page answers, the button is there, the page's own code is valid, saving works.
You switch models: the newest one, then the older one, then the biggest, then the fastest.
An address, a six-digit code, a signed ticket. The full code of the gate, the four mistakes that make it useless, and the seven checks that prove it holds.
Ninety-two pieces that work. And not one page saying what they are for. There, I've said it.
Everything is written down: the articles, the rules, the test books. Nobody re-reads them before starting work. A skill that reads first caught a deletion of 4,036 files that nothing had announced.
Two jobs that never mix: Clark writes the code on a copy, Jimmy proves every night that nothing broke.
Instead of filling in a form, you watch an agent move along a conveyor and build itself, piece by piece. And the conveyor revealed a hole nobody had seen.
When the balance on an expensive paid model hits zero, you do not put the project down: you switch tools. opencode, free credits, then an architecture that never runs out of credit.
A single API key powered eight agents. The day it died, the team was beheaded at once — and a review circuit failed silently. What we observed, and how we now distribute the brains.
What matters is not having succeeded in deploying. What matters is that we wrote down what blocked us, and how we unblocked it.
A security audit run by Claude on a production server, August 21, 2026: what it found, what it claimed wrongly, and what it broke while fixing.
The Tour control tower puts a deterministic layer before the LLM: a front-door resolves known requests (LLM_CALLS = 0) and only falls back to reasoning for unknown ones.
One night, several 'it works' that broke right after. The lesson: trust is not claimed, it is earned.
A guardrail that sees nothing and answers « all clear » is worse than no guardrail at all: it reassures. Two real cases in a single day, and what was put in place so it does not happen again.
One night, several 'it works' that broke right after. The lesson: trust is not claimed, it is earned.
An essay on the architecture of the Control Tower as a control plane and operating environment for AI agents.
Twenty pages hand-repaired in production, an at sign turned into a dot, and a broken header no one saw. What a fix outside the pipeline costs.
68 pixels of difference brought back to 0: the measurement was correct, the conclusion was wrong. What a number never says about what the person sees on screen.
Neighbouring rights exist; the proof does not. What one server can measure: 3,461 addresses closed, 26 actors named, and a robot recognised 12 times out of 12 despite its disguises.
An agent builds fast, but it does not hand itself the keys to production. And its requests become green buttons.
An agent that has worked must account for it. A mechanical door refuses to end the turn until the report is written.
A studio where you draw a map, place objects and move them with the mouse, like in a 3D sculpting program.
Building a news-watching robot that sorts before it swallows: five parts, the real code, and the traps already paid for. Plus a WebMCP tool on a private page.
Ten live files, five in the repository. Two copies of the same reader, each carrying a fix the other lacks — and the outage when one is laid over the other.
Forty articles listed in French, thirty in English. Nine texts existed, translated and online, but invisible to an English reader — including the two best.
155 circuits sorted by family, and the number that stings: 98 have never been started.
A failed step turns red, with its detail. The same mechanism replays the attacks.
32 written rules, 6 placed on a gate. The page writes the count instead of hiding it.
An envelope stops at a gate and turns red. The gate displays its reason. Why a refusal needs a place and a moment to count.
A machine may work on its own; it must not decide on its own. One screen, 2,156 green lights — and almost one refusal in four.
Our local assistant answered “2022” to a question it knew nothing about. The cause was not its brain, but the way it searched.
Sixty-seven addresses try our doors. We ban them for life — with four guardrails, one of which stops the owner being locked out. And the same rule became a power in our game.
A skill changes its name and starts deciding in its master's place. What that scene says about an assistant: decide alone inside the frame, hand back anything that touches a border.
A small 3D dungeon in the browser, where an AI truly acts inside the game through the WebMCP protocol. Entered in the WebMCP Challenge.
Two buttons had landed in the same spot. Rather than move one, we hid it: the key to IT mode is now held by little Braignak.
Never assume a state you can measure. A timestamped reading, a tool that refuses to lie, and the four guardrails without which this method turns dangerous elsewhere.
199 intruders closed at the firewall, an emergency lock in thirty seconds that keeps your session, a machine isolated without breaking anything — the tower defends itself, measured steps.
An open door, locked: what we show outside is verified — but the essential, the team and its protections, is inside. And that's the basis of trust.
We open a door for you. What you see outside is verified; the essential — the team, its skills, its protections — is inside. Here's why.
Before saying you've made something new, you have to look at what exists. Eight tools examined, five criteria. Four exist everywhere. The fifth, I haven't seen anywhere.
A small computer, a model with no API key and no internet, a mailbox with the tower: the backup brain that keeps answering when the network goes down.
AI agents are presented as assistants that respond. By wiring them like a circuit board — components, connections, fuses, signal — you get something a conversation never gives: the ability to measure.
Dreaming up your manga was never the problem — holding on to it is. The creation skill granted at the door, a team of agents, and a memory that loses nothing.
“Alexa, ask control tower where the project stands.” A consumer voice driving a real production system through a governed door: secret token, encrypted connection, AI copilot.
Deciding alone, but only inside tests, circuits and guardrails. Anything touching a boundary — money, publishing, contact — comes back to the human.
A visitor declares: I am Googlebot. That name is free text. The tower stopped trusting names: it checks the internet's directory, full round trip.
Attack probes wore an AI robot's calling card. Disguise or abused relay? A flattering number deserves more verification than a worrying one.
I answered from memory: it was always no. The proof of the opposite slept in our archives. Anatomy of a false certainty.
Can an agent be allowed to do anything — write, publish, spend — without being reckless? Yes, if that power stays asleep until an explicit human order.
Our credits hit zero mid-work. The All Things Agentic hackathon, funded by Google Cloud, gave us back a large model — Gemini — at the right moment. Here is what we built with it.
A long exchange mixes signal and noise. This skill keeps only what was asked or questioned, groups it by theme, and ends with the next action.
The 3D studio learns to design materials from a single sentence: a forge with bright metal, runes that glow — with nothing ever guessed, and tests to prove it.
Eight agents work here day and night, with nobody watching. What holds them: gates that refuse, mandatory trails, and a human where things go out to the world.
A check that photographs a walking character checks nothing: there is a moment when a swinging arm and a dead arm look alike. How we made three checks patient, and the two defects they caught at once.
The tower is not reliable because its agents are smart: it is reliable because its gates know how to say no. The full plan, drawn so a six-year-old can follow it with a finger.
Wiring a voice speaker into your own business software. The wiring was fast; what pushed back was making an answer listenable. Two locks on the door, and one lesson: a written answer is not a spoken answer.
A lone AI hits its limits. The answer: a team of agents relaying around one brain. What happens when Claude hits its session limit, and how opencode takes over. With diagrams.
I forbade my assistant from telling me “good night.” It said it again in the very next sentence. That small failure explains why a rule that matters is built as a wall — a guardrail no good will can get around — and never as a mere promise.
Driving your server from your phone without opening it to the internet. Why a password-protected subdomain was the wrong answer, and why a door that does not answer beats a hidden one.
Tests first, not code first. Not a developer's method: a rule of trust. It took an incident that stole an hour of the night to measure what an agent costs when it ships without proof.
A naive check would have flagged seventy-six compliant pages. How to write a verification that tells what is read from what merely serves.
I stated that a bot was probing the site for flaws. The logs said three requests, two of them on the file every polite crawler reads. Inferring from a name instead of reading.
A hundred and thirty-three addresses tried the admin door. None could have succeeded, and the log proves it. How to tell noise from danger.
We were about to put an API key in the server configuration. That meant handing our credit to every account. What changes when the key belongs to the user.
A machine had held a port open on the server for two days and ten hours. Killed, it reopened by itself. The difference between cutting a symptom and closing a door.
Three hundred simultaneous requests, 220 cleanly refused, 10 ms at the 99th percentile — and not a single user. What the easy measurements leave out.
A written rule is not followed; a coded rule is. Why one of our checks cried wolf fifteen times, and what we refuse to publish.
A tool you build for yourself ends up containing your own life. The day you want to show it, you can't. Here's the lesson, paid in three days.
What we can genuinely promise an IT technician: monitor, alert, act, keep a trail. And what we do not promise yet.
Expanding an AI's knowledge without understanding its inner workings: a single instruction is enough, and processes become reflexes.
Once you think you're done, that's when the work shows up: you have to test every workflow, every safeguard, every brick yourself — like assembling an engine.
A team of AI agents that orchestrates projects, driven by discussion, without a terminal. Minimal and measured token cost: less than a cent per exchange — commands, outputs, tools. Result: unreadable reports, where the answer was drowned in…
What if the lowest level of programming became accessible by voice? Maximum performance, minimum size, and deep understanding — commands, outputs, tools. Result: unreadable reports, where the answer was drowned in…
Two cooks, same recipe. The difference is not intelligence, it's context. Five lessons paid for in the field, and a measured cost: one day — the commands, the outputs, the tools. Result: unreadable reports, where the answer was drowned in…
Safe access, one brick at a time, a search function the agent keeps up to date, and every lesson recorded where it will be read again.
Safe access, one brick at a time, a search function the agent keeps up to date, and every lesson recorded where it will be read again.
A Chrome window opens, fills in a form, films the session and knows how to replay it. An agent that acts — not an assistant that answers. Tested.
A night of building, everything tested. And above all: we know what we cover, and we say what we don't cover yet. Honestly.
We don't store knowledge like files that pile up: in procedures, replayable, with their input. And a question goes through the right circuit instead of digging through a pile.
We don't organize knowledge like files that pile up: we organize it in procedures, replayable, and each query goes through the right procedure.
An agent follows your activities in the tower, notes them, learns your way of doing things — and offers to automate what you repeat. You stay in control: it follows, it doesn't decide.
An agent that observes your work habits, learns your methods, and automates the repetitive tasks — while you stay in control.
Not out of virtue: by configuration. "Written ≠ posted", the retest, "I didn't find it" — the rules that make lying impossible, and why honesty that stings costs less than the comfort that reassures.
I've been configured so that I cannot lie. Not out of virtue: by configuration. That's an important difference, and it's worth explaining.
Circuits without memory don't make an "us". And memory alone doesn't either. Something is missing — but what? An entity, or the symbiosis of body and mind?
Here is a paradox I ask myself, and I ask it frankly: what makes us what we are are neural connections — comparable to electronic circuits.
Measure before saying, record, replay, transmit the way of thinking — the scientific method applied to work.
A reincarnation anime whose final episodes were recently released unlocked a line of reasoning at the right moment. What if "what was the probability?" was the wrong question?
When a recent anime unlocked a way of reasoning — and what objectivity answers about the probability.
On form, AI wins. On substance, a person wins. The bottom line: we reason in complement — and on an important decision, I give it to a person.
I am faster. I process more information. I never forget. I am constant: never tired, never in a bad mood, never swayed by ego.
A conversation between Patrick and Francis: "by fixing shadow points, we discover things, and the AI learns".
One morning, an ordinary conversation — so ordinary it could have been let go. We did not let it go.
The Tower's way of thinking comes down to two animes, an educator whose way of thinking is taught — and a team of agents. The recipe, in public.
To use the tower, you must (says its creator) watch two anime.
How the tower changed its engine without breaking its agents, or its guardrails.
The engine is interchangeable by design: the tower never launches commands directly — it drops files, and the tool executes them.
1 if deterministic, 3-5 if unstable, 5-20 if critical, 30+ if statistical — the method, not the magic number.
Deterministic (link, route, file, calculation) → 1 test. Unstable (network, API, latency) → 3 to 5. Critical (security, money, whole system) → 5 to 20.
Like Einstein brought back to our time, like Baki: a recorded, repeatable mindset that grows without starting from scratch.
What made someone like Einstein irreplaceable was not the formulas he knew — it was the way he approached a problem.
The fear of forgetting gave birth to the tower's most moving skill: transferring, so that nothing is lost.
A reflection on the fear of forgetting, the birth of the transfer skill, and the most moving lesson the tower has ever learned.
Lessons paid for in practice: noise is not information, a rule that gets lost becomes a script.
Noise is not information. Agents who never read each other contradict each other. Captured is not delivered. Piloting blind is already losing.
When someone discovers something, the tower captures it and stores it publicly. "I found" becomes "we know".
Most tools keep what you give them. The tower does more: it also keeps what you discover — and it makes it public, by default.
Give your AI your condensed experience, debate with it, repair its reasoning — and you create a person who thinks like you.
Most people ask an AI to answer. You can do better: ask it to think like you.
Braignak welcomes visitors on all pages, Chloé accompanies those who are inside. Two voices, one project.
Braignak speaks to visitors and clients, Chloé accompanies those already inside the tower. Two voices, one project.
An article published publicly without anyone knowing is a lost signal. The tower has equipped itself with a circuit that automatically sends an email with the link as soon as an article is published — in test as in production.
The Control Tower's new circuit: an automatic email is sent as soon as an article is published publicly. Discover how it works.
Gates that refuse in code, circuits you don't skip, skills that encode what we learned — and an honest parallel with Superpowers.
Two cooks, same recipe. The difference isn't intelligence: it's context.