AI loops are becoming one of the more important ideas in agentic AI: systems where one agent can prompt, evaluate, or redirect another agent until work improves or finishes.
At Meta’s @Scale conference on Friday, Claude Code creator Boris Cherny was asked whether loops are only another hype cycle or something real. His answer was clear: “yes, they’re for real,” he said.
What AI loops change
Cherny described the shift as part of a broader evolution in software work. “Two years ago, we wrote source code by hand. 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,” he continued.
That framing matters because it moves AI coding beyond a single agent responding to a single prompt. In a loop, the work can continue through repeated agent activity. The system does not simply wait for a human to ask the next question; it can keep checking, suggesting, and improving.
Cherny argued that this is a major step. “As big as the step from source code to agents was, loops are just as important and as big a step.”
In his own workflow, Cherny said he runs loops that keep watching a changing code base. One agent looks for ways to improve the code architecture. Another searches for duplicated abstractions that could be unified. Those agents submit pull requests like other coders, and because the code keeps changing, the loops keep running.
From managing agents to supervising systems
Most agentic AI work still depends on careful human management. A user sets a goal, checks progress, and tries to keep the agent from drifting away from the task. That makes the agent useful, but it also keeps the workflow bounded.
AI loops expand that model. Instead of asking an agent to complete one discrete unit of work, a user can authorize agents to operate continuously in the background. The goal is not only task completion, but ongoing improvement.
That creates a different kind of trust problem. A user is no longer only trusting an AI model to answer a question or write a block of code. The user is trusting a system of agents to keep acting, reviewing, and proposing changes while the underlying project continues to evolve.
The appeal is obvious in software development. Code architecture can always be examined again. Duplicate abstractions can appear as a code base grows. Pull requests give the loop a familiar channel for review, which means AI work can enter the same process as human work rather than bypassing it.
Why the idea feels familiar
Loops are not a new concept in computing. Recursive loops, where a function calls itself to repeat an action until a stopping condition is met, are a basic computer science idea.
Agentic loops work differently because the stopping logic can be non-deterministic. Instead of a clear condition ending the process, a sub-agent may decide whether the goal has been met. That makes the loop less predictable than a classic programmatic loop, but the underlying pattern is still repetition toward an outcome.
One example is the Ralph Loop, named for Ralph Wiggum. The method summarizes what the model has done and asks whether the goal has been accomplished. It is meant to help when AI models lose track during longer runs by bouncing the process back until the task is complete.
This is why AI loops fit naturally into the agentic AI trend. Once AI systems began completing multi-step tasks, it was logical that developers would experiment with AI overseeing AI. The loop gives that pattern a simple structure.
The compute tradeoff
Another way to understand AI loops is through the push for more test-time compute. OpenAI researcher Noam Brown observed earlier this month that contemporary models can solve nearly any problem if enough compute is applied.
For a loop, that means persistence becomes part of the strategy. If the system has not solved the problem yet, it can keep spending compute on the task. For hill-climbing problems such as improving a code base, the model can make incremental improvements until it reaches a threshold.
In Cherny’s example, the process can continue as long as there is compute available. That can be powerful, but it also changes the economics of using AI.
AI loops burn through tokens faster than simple Q&A chatbots. Because the point is to keep the loop running, the cost does not have a natural ceiling. The source notes that this may be acceptable for Anthropic, which is in the token-selling business, but it can become expensive for other users.
Where loops may be useful
The strongest case for AI loops is work that benefits from repeated inspection and incremental improvement. A code base is a natural example because architecture, duplication, and pull requests already fit a review cycle.
But the setup matters. The loop needs oversight of token spend, drift, and other classic AI issues. Without that, the same persistence that makes loops useful can also make them costly or hard to control.
The practical question is not whether loops can run. They can. The question is whether the value of continuous background work is high enough to justify the compute and supervision required.
If the answer is yes, AI loops could become a major part of real work with agents. They turn agentic AI from a tool that waits for prompts into a system that keeps looking for the next useful change.