Boris Cherny, the creator of Claude Code, just made one of the boldest claims yet about where AI work is headed. Speaking at Meta’s @Scale conference on Friday, he was asked point blank whether “loops” are the next hype cycle or the real deal. According to TechCrunch AI, his answer was an emphatic “Yes, they’re for real.”
What stands out here is how he framed the size of the shift. “Two years ago, we wrote source code by hand,” Cherny said. “We started to transition so agents write the code. And now we’re transitioning to the point where agents are prompting agents that then write the code.” In his words, that jump is “just as important and as big a step” as the move from hand-written code to agents.
What a loop actually is
Strip away the buzz and a loop is simple. Cherny runs agents that never stop. One keeps hunting for ways to improve his code architecture. Another looks for duplicated abstractions it can merge. They open pull requests like any other engineer, and because the codebase is always changing, they just keep going.
TechCrunch AI makes a smart point: this isn’t brand new. Recursive loops, functions that call themselves until a condition stops them, are intro computer science. The twist is that a subagent now decides when to stop instead of a fixed rule. One popular version, the Ralph Loop (yes, named after Ralph Wiggum), just keeps asking the model “did you finish the goal yet?” until the answer is yes. It’s a crude trick to keep a model from getting lost on long tasks, and it works.
Why it matters now
Loops are really a bet on test-time compute. OpenAI researcher Noam Brown noted this month that modern models can solve almost any problem if you throw enough compute at it. Loops take that literally: keep spending compute until the job is done.
That fits a specific kind of work perfectly. Hill-climbing problems, like steadily improving a codebase, let a model make small wins again and again until it clears a threshold. There’s no single “done.” There’s just better.
Here’s the catch, and it’s a big one. Loops burn tokens fast, much faster than a normal chatbot exchange. Since the whole point is to run nonstop, there is no ceiling on spend. TechCrunch AI is blunt about the conflict of interest: that’s great for Anthropic, which sells tokens for a living. For everyone else, it can get expensive in a hurry.
The competitive read
This is significant because it reframes what “good AI usage” looks like. The advice for the past year was to babysit your agents: set clear goals, check discrete progress, don’t let them wander. Loops flip that. You’re handing a swarm of agents a standing order to work in the background, forever. That’s a lot of trust to place in a system that still drifts.
It also tells you where the frontier labs want the money to flow. More autonomy means more compute, and more compute means more revenue for whoever sells the tokens. When the guy behind Claude Code calls loops a generational step, read it as both a genuine technical conviction and a business direction.
Practical takeaways
If you build with or buy AI, a few moves make sense right now:
- Pick the right problems. Loops shine on open-ended, incremental work (refactoring, test coverage, optimization), not one-shot answers. Don’t loop a task that has a clear finish line.
- Cap the spend before you start. Set token budgets and hard stops. “No ceiling” is a feature for the seller, not for your invoice.
- Build oversight in. Watch for drift, review the pull requests, and keep a human gate on anything that ships.
- Run a small pilot first. Measure cost per useful change against doing it manually. If the math doesn’t beat a junior engineer’s hour, wait.
The honest read: loops can deliver staggering results on the right problem with the right guardrails. They can also quietly drain a budget on the wrong one. The teams that win will be the ones who treat compute like cash, not like a faucet they forgot to turn off.
For the full talk and Cherny’s specifics, the original report at TechCrunch AI is worth your time.