AI Agents and Engineers: 700 Files per Commit and the Hidden Cost

From 700 files in one commit to the question of who pays for tokens

Four hundred changes a day for three consecutive weeks. A final deployment of 700 files in a single commit. The project appears finished in four days, colleagues no longer recognize the code, and nobody knows if it will crash when it gains visibility. The mechanism is always the same: requirement files in plain text, and the agent implements, readjusts, and tries again. The promise that artificial general intelligence will eliminate computer engineers in six months already has its first real test field, and the result is neither a mass layoff nor a clean revolution.

The 700-file commit that no one can read anymore

An engineer went on vacation for three weeks and upon returning found 160 commits and 70% of the code changed. No one could explain it. Interrogating team members separately revealed the pattern: someone had used agents to adapt a small change to a new backend method. He's been reverting for three and a half weeks.

The issue has two interpretations. The generous one says that if a problem is broken down into 200 pieces, the AI solves them well, and the pure backend—without interruptions or quirks—holds up. The other warns that most of these changes miccionan nothing: cleaning up logs, touching every log call, hardening validations. Lots of action, little business. The payer cares about functionalities and cost, not the tidiness of the history.

Programming in ten minutes and testing in two days

Here's the data that fuels the hype: what took two months two years ago is now done in ten minutes. The fine print appears right after. What used to be validated in ten minutes now requires two days, and in critical systems—banking, aviation, transport, medicine—that delay isn't a calendar inconvenience, it's an operational risk. The code comes out fast; confidence does not.

There are documented experiences of over 60,000 lines of code generated in a month and a half, with robust architecture and technologies the author hadn't touched before, on work that a senior would have signed off in a year and a half. The conclusion drawn by that same profile is uncomfortable: they wouldn't recommend anyone study computer science right now. It's another matter whether the system can pass an audit with the server shutting down mid-process.

Who pays for the tokens in all this?

The question runs through the entire conversation, and almost no one answers it. Four hundred files a day burn a lot of context window, and the cost per token scales with volume. Twenty-euro subscriptions hit a five-hour usage limit that leaves tasks unfinished; free models are discarded as insufficient; the best ones are more expensive. The company pays, obviously. Until someone looks at the bill.

Added to this is a fundamental problem that more computing power won't fix: if the generated value can be replicated by the client themselves, where does the margin lie? And if everyone consumes models without producing new data, who feeds the next generation?

Vacant position, amortized position

The impact is already being measured in payrolls. From an IBEX company comes the testimony that, for months now, any position that becomes vacant is amortized with AI functions. The virtues of an automated call center have also been shown, with hundreds of jobs heading for the exit in a matter of days. Junior profiles are the first to feel the ground shifting.

The medium-term consequence being considered is more nuanced than it appears: if one person produces for ten, demands will rise by a hundred, and ten will be hired again. The next two years, however, look bad. And the failures, few but very significant.

The bubble that finances the circus

The most cynical and solid argument: the tool works, but today it's not profitable. Chip manufacturers know this, they finance the entry and create a perverse race where those who don't join the circus see their competitor devour them. A bubble, yes. A correction of expectations and prices, also. With Google and Microsoft having the muscle to lick their wounds and many others without it.

With these elements, the disappearance of the computer engineer in six months is far from what the data supports. Productivity rises, cost rises, failures become rare and enormous, and no one has managed to put a price on the disorder. In six months, say the skeptics, we'll still be rowing the same way. It's exactly what was said six months ago.