Anthropic recently published a piece titled "When AI builds itself," and it’s one of those rare long-form articles worth the read. In the paper, Anthropic shares that AI has begun accelerating its own development to the point where more than 80% of the code merged into Anthropic's codebase is now written by Claude, and the average engineer there ships roughly 8 times as much code per quarter as they did 2 years ago.
Anthropic’s argument carries weight not just because of their scale, but because the company is already operating in the conditions it’s describing: systems contributing meaningfully to their own iteration cycles, and engineering output compounding as a result. The authors are honest about where these AI-driven builds could lead, including a future in which systems design their own successors with little human involvement. Moreover, they make a careful case that the world should at least have the option to slow down so that safety work and institutions can catch up.
It’s a serious piece by serious people, and we think they are right to raise the question: What happens when AI builds itself?
There’s a point that the story repeats, and one we keep coming back to ourselves: Even in Anthropic's own telling, the part of the work the machines have not taken over is judgment. Humans still need to decide which problems are worth solving, which results are trustworthy, and when an approach has become a dead end. They call it research taste, and they are candid that it is the one thing their fastest models still cannot do reliably.
We read that and thought, “that's the whole story.” The part that remains human is not a leftover, but the part that was always the hardest to develop and the most valuable to have.
This is what we have believed at FullStack from the beginning. For more than a decade, we have been building a technology talent ecosystem across Latin America, and the thing we are proudest of is not the number of engineers we can place, but the kind of engineers we help create.
We look for people with real curiosity and genuine range, we put them through a vetting process that is honest about both their strengths and their gaps, and then we keep investing in them long after any contract is signed. We mentor them, we hand them hard problems before they feel fully ready, and we give them the room to grow into roles they could not have stepped into a year earlier. The result is a community of engineers across the region who not only know their frameworks, but who know how to learn the next one quickly, and who can sit inside a customer's business and work out what actually matters.
That last skill is the one that makes or breaks an engineering team. The engineers who thrive aren't the ones who memorized yesterday's stack, but the ones who can pick up whatever is in front of them and bend it toward a real outcome. We have watched our teams do this through wave after wave of change, from cloud to mobile to data, and now to AI. Each and every time, the same pattern holds: the tool matters less than the person wielding it. A capable engineer with a new tool is a force multiplier. A new tool without a capable engineer is just expensive software.
AI does not break this pattern. If anything, it makes it sharper, because the distance between someone who can direct these systems well and someone who cannot is now enormous. So we are doing exactly what we have always done, now with our sights set on AI. We are training our engineers to work alongside these systems rather than around them, to treat a capable model as the fastest junior teammate they have ever had, and to supply the judgment that the machine still lacks.
We are teaching people to ask the right question before they automate the answer, to review what an agent produces with a skeptical and experienced eye, and to own the outcome rather than just the keystrokes. We embed these people directly with our customers, close to the work and the decisions, because that is where taste and context actually live. The Anthropic piece worries that human review could become the bottleneck as AI speeds everything else up, and we agree. That’s exactly why we invest so heavily in the people who do that reviewing and directing.
So when we read a call to slow down and let people catch up, we don’t hear a threat to our model. We hear our model described back to us. The long-term success of AI will not be decided by how fast it can build itself, but by whether there are enough capable, grounded, well-supported people to point it somewhere worth going.
We have spent years finding and building exactly that kind of engineer across Latin America, and we are going to keep building them. That is our bet, and we are more confident in it today than ever.
.svg)
