AGI Without Developers Fails: 700 Files Per Commit

AGI was supposed to replace developers, but now shows 700 files per commit

The prediction repeats with clockwork precision: general intelligence will eliminate the need for software engineers in six months. The argument is seductive. Anyone can ask the model for the software they need — "I'll make an SAP for my 1,000-employee company," summarizes the current trend — but economist Benedict Evans debunked this long ago: it confuses how most people think with where software really comes from.

What has peine in real projects in recent days points to something else. Not that AI does not work. But that the bottleneck has moved.

Why is testing code more expensive than writing it?

Because no one fully knows what the machine generated. A fact circulating among developers summarizes it: what used to take 10 minutes to test now takes 2 days. AI writes in minutes, but validating that code has become the expensive part of the process.

The effect multiplies in critical systems. Banking, aviation, transport, medicine. There the margin narrows: either test thoroughly, or face the day the service fails for weeks. The paradox is uncomfortable. The same engine that reduces writing time multiplies verification time, and those who used to review now only sign off.

700 files in one commit and 400 changes per day

The most repeated case has a scandalous number. A platform engineer had been generating between 300 and 400 daily changes for three or four weeks, with peaks of a commit of 700 files, based on requirement files the model implemented alone. The rest of the team lost knowledge of the code and was left out of their own project, watching a program grow that they could no longer read. The appearance was impeccable in four days. What will happen when the system gains visibility, no one signs off.

It is not an isolated case. Another team returned from three weeks of vacation with 160 commits and 70% of the code changed. Three and a half weeks reverting and patching. On the opposite extreme, someone built a robust architecture on a server of more than 60,000 lines and the model implemented it "to the letter": a month and a half for what would have been a year and a half of senior work, with technologies they had never touched before.

The entire spectrum depends on who sets the requirements.

From agile development to specification: the method changes overnight

Some teams have switched from agile to specification-driven development (SDD) overnight. The logic: narrow the field so the model does not lose context. Those who have tried it warn of the usual: current AI has a context that is too general and "forces you to choose: either AI or me, the two cannot coexist."

Behind this is the bill. How much does AI cost per year? Who will generate new data when we are all consumers? Some already pay 100 euros monthly between frontier model and APIs. And there are 3 trillion dollars in unfunded spending commitments attributed to the sector, with the warning that the token price of closed models will collapse to the levels of open models. More reason to look at local models or alternatives for compliance reasons, such as using a French AI instead of American ones.

Local AI that works and that that produces nonsense

Not all is expensive smoke. Models running locally work "reasonably well for well-defined problems": string handling, persistence of structures in static memory, extensions on an already prepared architecture. The phrase that orders the debate comes from the same place: "do not ask it to make me a program for this, because it will make a mess."

From this arises the conclusion that is gaining ground. The engineer does not disappear; they shift toward integrating pieces and defining what is wanted. Software becomes integration, not writing from scratch. The boundary between what a model resolves alone and what requires a human narrows each month, but does not disappear. They say this even those who would have laughed before: given the warning that junior profiles will have a hard time, someone concedes that maybe we need to look for them again, even if only for user testing.

What if the problem ends up being that no one is needed? The most apocalyptic current calls for basic incomes to hold out until demographics do the work, with birth rates plummeting halfway around the world. Sounds like a movie. No one has presented the spreadsheet.

With these differentials, the replacement of the developer should be imminent. No one sees it coming. The usual skeptic had it noted: "If there is no money bet, I will not insist on this. In 6 months we will still be rowing as always." And they were told that this was said six months ago. There, exactly, the calculation gets stuck.