@BlueRockTEL Une solution de facturation et CRM complète qui vous accompagne dans l'intégralité de votre process de billing Telecom et cloud, de l'intégration des CDR à la génération de vos fichiers de prélèvements SEPA. Et bien plus à découvrir : https://t.co/r87819DA4g
Everyone is building AI coding agents.
I'm more interested in building AI development pipelines.
My workflow looks like this:
• I write the initial brief.
• The AI decomposes it into tasks and proposes acceptance criteria.
• I review and approve both.
• The AI implements the task.
• It validates its own work against the approved criteria.
• If validation fails, it retries.
• After the third failure, it is forced to change strategy.
• When finished, it proposes follow-up tasks across related projects.
The entire system is written in 100% Bash.
No orchestration framework. No agent platform.
Just shell scripts.
https://t.co/bJvD6sEFVm
If you have to watch Claude work, you don't have an agent.
You have a chat window with tools.
My rule for AI agents is simple:
I am not in the loop. I am at the gate.
I approve the task and its acceptance criteria. Then I leave.
The agent either completes the job, or gives it back after three failed attempts.
Human attention is the expensive part. Not tokens.
A monitored 3CX host goes down. Your team should never learn about it from an angry customer.
The alert already fired in seconds. It just landed in an inbox nobody was watching at 08:14. Here's how you close that gap.
You cannot redistribute a thought.
On 3 November, California asks its voters to take 5% of every resident billionaire's net worth. Proponents promise $100 billion for health care.
The idea did not start in Sacramento. Gabriel Zucman, among others, was inescapable in the French media a few months ago, making precisely this case. So the same argument has now reached me twice in one year, in two languages.
The number is not the interesting part. The premise underneath it is: that the $100 billion already exists, in a pile, somewhere, waiting to be moved.
It doesn't.
Ayn Rand noticed something about Americans that almost nobody notices. Everywhere else, people say they earn money, they get money, they take money. Americans were the first to say they make money.
That verb is not an accident of vocabulary, it is a confession: wealth is manufactured. Not found lying on the ground. Not lifted off someone else's table. A civilisation got built on that sentence.
I think of the economy as one gigantic flow of information, spreading out over time. Not a pot. Not a pie. Not a cake waiting to be sliced by whoever holds the knife.
A flow. What produces it is human intelligence, applied to matter, and then traded.
Voluntary exchange is where the flow starts, and it is embarrassingly simple. I hand you €30 for a bottle of wine I have never tasted, because I want the bottle more than I want the €30. You hand me the bottle, because you want the €30 more than you want the bottle. Five seconds later we are both richer than we were, and nothing was taken from anyone. Net contribution: doubly positive.
That is growth. Not a statistic. Not a quarterly print. Billions of tiny doubly positive events, every day, everywhere, most of them too small to notice.
Now reverse it.
Force the exchange. Take instead of trade. One winner, one loser, nothing created, nothing doubly positive anywhere in it. Zero sum is not the economy. Zero sum is what coercion does to the economy.
Socialism starts from the opposite premise: the wealth is fixed, the pile already exists, and the only real question is who holds it. Follow that premise honestly and you inherit its whole moral vocabulary. The merchant is a thief. The employer is an exploiter. The investor is a parasite.
Those are not insults thrown in anger. They are logical consequences of a false axiom, and once you accept the axiom they are perfectly coherent.
The difficulty for the axiom is reality, and reality is embarrassing. Eight billion people trading daily. Extreme poverty down from around 40% to under 10% in a single lifetime. More wealth, and more people in work, than at any moment in human history.
If all of that was stolen, stolen from whom? The dead?
There are only two replies available.
The first is to deny the numbers. A lie, an accounting trick, GDP propaganda: choose your version. Then look at a Nigerian teenager holding a smartphone, and at a Chinese grandmother who remembers famine, and tell them nothing happened.
The second reply is cleverer, and it is now the dominant one: concede the growth and relocate the theft. We did not steal from each other, we stole from the planet. Humanity as a collective looter of finite resources, running down a stock it never created. Same arithmetic, new colour. This one is green.
It makes the identical error, one layer down. Sand sat under our feet for 300,000 years and was worth nothing. Oil was a nuisance that ruined good farmland. Uranium was a rock. Then someone had an idea, and only then did they have value.
Value is not in the resource, but in the intelligence applied to it. Which is why the doom maths never works: it counts atoms and forgets ideas. Atoms are finite. Ideas are not. We have never run out of ideas, and that is the only reason we have never run out of anything else.
So tax the fortune if the voters want to. That is a political choice, it belongs to them, and it will move some money once.
But understand what you cannot do. You cannot tax the thought that made it. You cannot vote a new one into existence.
Wealth is not found. It is not seized.
It is made.
If you have to watch Claude work, you do not have an agent.
You have a chat window with tools. The industry now sells that as "agentic". It is not.
An agent does the job when you are not in the room. Everything else is a demo that needs an audience.
I run a telecom SaaS. My best hours belong to the core product. The rest of the work is real, necessary, and not where the company lives or dies.
Migrations. Housekeeping. Specs that must become code. That is the work I refuse to sit through.
So I built a pipeline around Claude Code. Not a product. Bash scripts. Markdown files in folders. Cron on a Linux machine.
The idea is almost boring when you write it down.
A rough idea lands in briefs/. Claude rewrites it into a structured task with acceptance criteria. That prepared file waits in a review folder. I read it for about two minutes. If the plan is sane, I move it to inbox/.
Cron picks it up. Claude runs headlessly in tmux, grades its own result against the criteria I already approved, and retries when something fails. After three failures it stops and asks for a human. After a pass, the result sits in done/ and a desktop notification fires.
I am not in the loop. I am at the gate.
That gate is the whole design. The model is cheap relative to attention. Attention is the scarce thing.
A pipeline that never asks you to review is how unattended software quietly wrecks a working tree. A "workflow" that still needs you as the permanent audience is just pair programming with better autocomplete.
There are only two ways most people use these tools.
The first is the live session. You type. You wait. You correct. You wait again.
The model is fast. You are the bottleneck. Your calendar is the ceiling.
The second pretends to be different and is not: more agents, more panes, same job. You become air-traffic control for a swarm that cannot land without you.
I want a third option. Work enters as text. Work leaves as text, plus the files that changed. Between those two states, the only human obligation is a short, deliberate review of the plan. Not a live audience for the typing.
What sits at the gate is the plan: the steps, the acceptance criteria, the shape of "done". What does not sit at the gate is the keystrokes.
If the criteria were wrong, that is on me. If the criteria were right and the run still failed three times, the task goes to a review pile and waits for a human brain, not for another blind retry.
The scripts are open source: https://t.co/EOSrQO15US. Clone them. Point them at a task tree. Drop a Markdown file in a folder. Walk away.
You do not need a better model to free your attention.
You need a queue that refuses to waste it.
The agent is not the chat.
It is the work that keeps moving when you leave the room.
Utiliser @sellsy_app comme CRM et solution de facturation, tout en "déléguant" la valorisation des usages (#Telecom mais aussi #Cloud et #Bureautique) à @BlueRockTEL ? C'est aujourd'hui une réalité : la facture "calculée" par BlueRockTEL est directement postée dans Sellsy via API
Le palmarès des 500 "Champions de la croissance 2021" a été publié par @LesEchosWeekEnd
unyc est heureux d’y être classé à la 266e place, parmi les entreprises technologiques les plus dynamiques en France.
Un grand bravo à nos équipes et un gros merci à nos partenaires & clients.
✨ In this thread I'll list tactics you can use to write cleaner code in Laravel.
As you use them repeatedly, you'll develop a sense for what's good code and what's bad code.
I'll also sprinkle some general Laravel code advice in between these tactics.
THREAD