Every December I write a small review of the year, mostly for myself. Writing forces me to be honest about what actually changed and what was just noise. And 2024 was a strange year, because the thing that changed the most was not a framework or a language. It was how we work.

AI assistants became furniture

In January, using an AI assistant at work still felt a bit like a confession. Some people loved it, some people rolled their eyes, and most teams had no clear policy. By December, it is furniture. Nobody asks anymore if you use it. The question became how you use it, and how much you trust what comes out.

For me the shift was practical, not emotional. I stopped writing boilerplate by hand. Test scaffolding, mapping code, small scripts to move data around, draft documentation. The assistant does the first 70 percent and I do the last 30. That last 30 is where the job lives now. My honest estimate is that it saves me a few hours a week. Not the 10x that the conference talks promise, but a few hours a week, every week, is real money when you multiply it across a team.

RAG went to production

This was also the year I saw retrieval augmented generation move from demos to production systems. The demo always works. You feed some documents, you ask a question, the answer looks magical. Then you put it in front of real users with real data, and you learn the hard lessons. Chunking matters. Stale documents poison answers. Users ask questions in ways your retrieval never expected. And evaluating quality is genuinely hard, much harder than building the pipeline itself.

The teams that did well were not the ones with the fanciest models. They were the ones that treated it like normal software engineering. Clear ownership of the data, monitoring, feedback loops, and someone responsible when the answer is wrong. The model was almost the easy part.

The agent hype started

In the second half of the year the conversation moved to agents. Systems that do not just answer, but plan and act in many steps. The demos are impressive and the hype is loud. I am curious but careful. Every step an agent takes is a step that can go wrong, and errors multiply across steps. I think 2025 will tell us if this is the real next thing or another overpromise. In late November, Anthropic announced something called the Model Context Protocol, a standard way to connect AI to tools and data, and it might quietly matter more than the louder demos.

The quality bill came due

Here is the uncomfortable part of my review. More code got written this year, and more of it needed review. AI writes plausible code, and plausible is dangerous. It compiles, it looks right, it passes a shallow review. Then three weeks later you find it handles the happy path only, or it duplicated logic that already existed two folders away.

I noticed this in myself too. When the code appears in seconds, there is a temptation to accept it in seconds. I had to build a new discipline, reading generated code with the same suspicion I would give to a pull request from a stranger. Robert Martin has been saying for years that our job is to be the last line of defense for quality. That sentence aged very well in 2024. The cost of writing code went down. The cost of bad code in production did not move at all. The gap between those two is risk, and someone has to carry it.

Fundamentals still won

If I look at what actually made projects succeed this year on my team, the list is boring. Clear requirements. Good domain boundaries. Tests that people trust. Honest conversations with stakeholders about scope. None of that was replaced by AI. If anything, AI made the fundamentals more valuable, because when code is cheap, knowing which code to write becomes the whole game.

I think about this from the business side a lot. A company does not pay us to produce tokens of code. It pays us to reduce risk and create value with technology. The tools changed this year. That contract did not.

What I am taking into 2025

Three small commitments for next year.

First, keep using the tools daily, but measure. Hours saved, bugs caused, real numbers, not gut feeling.

Second, get better at review. Reading code critically is becoming a bigger part of the job than writing it, and I want to be good at the part that grows.

Third, protect the fundamentals. Domain modeling, testing, talking to the business. The shiny stuff changes every six months. The boring stuff compounds.

When I moved from Brazil to Canada in 2017, I bet that the fundamentals of my craft would travel with me even when everything else changed. Seven years later, with AI changing everything else again, I am making the same bet. So far it keeps paying.

Happy holidays, and see you in 2025. It is going to be a loud year. I will try to stay quiet and useful.

Pax et bonum.