AI agents are now running AI agents in endless loops

Boris Cherny, the mind behind Claude Code, made a bold claim at Meta’s Scale conference last week. When asked if AI loops are just hype, he said they’re the real deal. Two years ago, humans wrote code by hand. Then AI started writing it for us. Now, we’re at the point where AI agents prompt other AI agents to do the writing.

Cherny shared how he uses loops in his own work. One agent constantly looks for ways to improve code architecture, while another hunts for duplicate code that can be streamlined. These agents submit changes like human coders and keep running as the code evolves. It’s a big leap from static prompts to a system where AI works continuously in the background.

The concept isn’t entirely new. Programmers have long used recursive loops where a function calls itself until a task is done. But AI loops are different. They rely on sub-agents to decide when to stop, not a fixed rule. Some loops, like the Ralph Loop, simply ask the AI if it’s finished its task, bouncing it back and forth until the job is complete. Others keep refining code endlessly as long as there’s computing power to burn.

This approach could be powerful. As models improve, they might solve problems by sheer persistence, throwing compute at a task until it’s done. But it’s expensive. These loops burn through tokens far faster than simple chatbots, with no natural limit to costs. For companies selling AI compute, that’s great. For everyone else, it might be a tough bill to swallow.

Cherny and others are betting that loops will become a standard way to work. We’ve gone from writing code ourselves to letting AI handle discrete tasks. Now, the next step is letting AI systems run themselves, with minimal human oversight.

Why does this matter to you? If loops take off, they could automate complex, ongoing work like optimizing software or managing systems. But they also raise questions about control and cost. What happens when an AI loop spins out of control or rack up massive bills? And can we trust AI to keep improving things without human checks?

The big unanswered question is whether the benefits will outweigh the risks and costs. We’re still early, but the tech is moving fast. Watch for real-world tests where companies try loops on meaningful problems, not just experiments.

Would you trust an AI loop to keep improving your work without you checking in?

Is the cost of endless AI compute worth the potential gains?

---

Filed under: AI, ArtificialIntelligence, AgenticAI, Coding, TechTrends