The AI world is getting ‘loopy’: Boris Cherny on why endless agent loops are the next big shift

At Meta’s @Scale conference on Friday, Claude Code creator Boris Cherny faced an unexpected question from the audience: “Are loops the next hype cycle, or are they for real?” Cherny’s answer was unequivocal. “Yes, they’re for real,” he said, drawing a direct line from hand-written source code to agents writing code, and now to a new paradigm where agents prompt other agents in continuous, self-sustaining loops.

What are AI loops and why do they matter?

Cherny described how, in his own workflow, one agent continuously searches for ways to improve code architecture while another hunts for duplicated abstractions to unify. These agents submit pull requests like any human coder, and because the codebase is constantly evolving, they never stop. The shift from managing discrete agent tasks to authorizing a persistent swarm of background agents represents a significant leap in trust and capability. For most users, the focus has been on setting clear goals and checking progress in discrete units. Loops remove that ceiling, allowing AI to work endlessly on open-ended problems.

Not entirely new, but fundamentally different

The concept of recursive loops — functions that call themselves until a condition is met — is a staple of introductory computer science. However, agentic loops operate on a non-deterministic logic: a sub-agent decides when to stop, rather than a hard-coded condition. This introduces both flexibility and unpredictability. One popular implementation, the Ralph Loop (named after The Simpsons’ Ralph Wiggum), simply asks the model to summarize its progress and check if the goal is accomplished. It’s a crude but effective way to prevent models from drifting off task during long-running operations.

Loops as a form of test-time compute

OpenAI researcher Noam Brown recently observed that contemporary models can solve nearly any problem if given enough compute. Loops embody this principle: instead of designing a perfect single-shot solution, developers can throw continuous compute at a problem until it converges. This is especially effective for hill-climbing tasks like codebase improvement, where incremental gains accumulate over time. Cherny’s example of an agent that keeps refining architecture indefinitely is a direct application of this idea.

The cost of endless loops

If this sounds expensive, it is. Agentic loops burn through tokens far faster than simple Q&A chatbots, and because the loop is designed to run continuously, there is no upper bound on compute spend. For Anthropic, which sells tokens, this is a feature. For enterprise users, it’s a cost that must be carefully managed. The benefits, however, could be staggering — particularly for code maintenance, security patching, and other tasks that benefit from relentless incremental improvement. The key will be building in oversight mechanisms for token spend, model drift, and output quality.

Conclusion

Cherny’s comments at @Scale signal that AI loops are moving from experimental trick to serious production technique. As models improve and costs evolve, the ability to run persistent, self-directing agent swarms could redefine how software is built and maintained. For now, the loop is still a nascent concept — but one that developers and executives alike should watch closely.

FAQs

Q1: What is an AI loop?
A: An AI loop is a configuration where an AI agent continuously runs a task — often prompting other agents — without a fixed endpoint, stopping only when a sub-agent or condition determines the goal is met.

Q2: How is an AI loop different from a standard agent?
A: Standard agents execute discrete tasks with clear start and end points. Loops authorize agents to work persistently in the background, making them suitable for ongoing improvement tasks like code refactoring or monitoring.

Q3: Are AI loops expensive to run?
A: Yes. Because loops run continuously, they can consume significantly more compute tokens than one-off queries. Costs must be carefully monitored, though the potential value for certain tasks can outweigh the expense.