Artificial intelligence, in practice
Artificial intelligence explained by someone who builds it.
Five text files in a folder. Copy all five from here, run it, and you have an agent that does marketing work — not one that talks about marketing.
Fabricio Costa
It was not an automated birthday card. She knew what year the company is living through and which part of it is mine — and I just…
Raising the wall got cheap. What got expensive is the foundation, the plumbing and a street address — and nobody asks for those i…
An agent is not a chat with a pretty prompt: it is four layers of files — and a fifth nobody puts in the spreadsheet, the cost. T…
Neither a workless utopia nor mass unemployment. The likeliest scenario is duller and more interesting: everyone becomes a manage…
Artificial intelligence is not replacing developers. It is raising the profession to a new level, where understanding problems, m…
An agent that dazzles in a demo and an agent that survives production are two different products. This is the map of what changes…
One agent solves a task. Several agents only solve a process if someone draws the org chart — including who reviews whom.
GPT, Claude, Gemini, Grok, DeepSeek, Qwen: stop asking which is best and start asking best for what.
When writing gets cheap, deciding gets expensive. The centre of gravity of the profession is moving from the editor to the decisi…
The capability gap shrank to the point where the choice became, for most cases, a decision about control and cost — not about qua…
Do not start with the technology. Start by writing the job description — including what it may not do without asking permission.
Not every new job has "AI" in the name. The ones growing fastest are the ones that got scarce precisely because AI raised the vol…
The tricks died. The discipline stayed — it just changed its name and its address: now it is called context engineering.
Free, subscription or API? The bill changes shape depending on who uses it, how many times a day, and whether you have a machine …
The risks that make the news are the cinematic ones. The risks that will hit you are dull, quiet and already happening.
AI did not hand a superpower to whoever was already big. It took the advantage of scale off the table — and that favours whoever …
The tool that makes you produce more can make you think less. You can keep the first effect without paying for the second.
Calling a model through an API is integration. Making an agent work on its own, with tools, memory and limits, is engineering — a…
The course that only delivers information is over — information became free and on demand. What is left of value is what AI does …
Both, and not in the same cities, nor in the same age brackets, nor at the same pace. That is where the problem lives.
Almost no AI project fails because of the model. It fails on the choice of problem, on messy data, and on people who were never b…
The right question is not whether AI writes better code than you. It is who answers when the code breaks at three in the morning.
One produces what you ask for. The other pursues a goal and decides the steps on its own. The distance between them is where all …
Assume the worst case for your ego: the machine writes better code than you, always. What still earns your salary?
Seven beliefs that get in the way of real decisions — from "it learns from you as you chat" to "the bigger the model, the better".
Who writes here
Senior software engineer · writes about applied AI
I have worked with software for two decades and I spend my days putting artificial intelligence to do real work. Here I tell you what works, what breaks and why.
Sobre mim e como as matérias são feitas · Projetos · Contato
