The man who built Claude Code has more or less automated his own job, and he is not shy about it. Boris Cherny, who runs Claude Code at Anthropic, says he does not really prompt the AI anymore. He writes loops that prompt it for him. “My job is to write loops,” he told the Acquired podcast, and per Fortune he has not handwritten a line of code in about eight months.
Angus Sewell (Angus the Nontechnical) picks this up and lands on the right instinct: do not just copy Cherny wholesale. Angus built his own self-improving loop, then deliberately left one human checkpoint switched on, keeping himself in the process on purpose rather than going fully hands-off. That restraint is the smartest part of the whole idea, and it is worth explaining why.
First, the plain-English version. A loop here is a script that keeps handing the AI tasks, reading the result, deciding what to do next, and going again, without you typing each step. Cherny reportedly runs hundreds of these agents at once, some days tens of thousands. It is genuinely a different way of working, and loop engineering has become the phrase for it.
Now the pushback, because there is a whopping survivorship bias hiding in the headline. Cherny is doing this inside Anthropic, on their models, their infrastructure and their safety rails, with the people who made the tool sitting next to him. The head of Claude Code fired himself is a great line and a terrible instruction manual for a solo founder running the same trick on a credit card. Fully autonomous loops fail in slow, expensive ways. They loop on a wrong assumption for six hours, burn tokens, and hand you a confident mess.
Which is exactly why Angus keeping one setting in human hands is the move, and I would go further. For most people the right loop is the one that does ten steps then stops and asks before anything irreversible, not the fully hands-off kind. Let it plan, draft, refactor and test on its own. Make it check in before it deletes, deploys, spends or emails. The value of a loop is the boring hours it saves you, and the danger is the one unsupervised decision that costs more than all those hours combined.
Copy the ambition, not the recklessness. Write loops, absolutely. Just build the off-switch first, and keep your hand near it.
Did you know: Anthropic has said the overwhelming majority of Claude Code’s own code is now written by Claude Code, which is either the future of software or a very elaborate way to make yourself redundant, depending on the day you ask.