At Meta's @Scale conference on Friday, Claude Code creator Boris Cherny faced an unexpected first question from the audience: “Are loops the next hype cycle — or are they for real?” His answer was emphatic: “Yes, they’re for real.”
Cherny explained that just two years ago, developers wrote source code by hand. Then agents began writing the code. Now, he said, we are transitioning to a world where agents prompt other agents to write the code. “As big as the step from source code to agents was, loops are just as important and as big a step,” he emphasized.
What Are AI Loops?
Cherny later detailed the loops he runs in his own work: one agent continuously looks for ways to improve code architecture, while another hunts for duplicated abstractions that can be unified. These agents submit pull requests like any human coder. Because the codebase is constantly changing, they never stop running.
In 2026, as agentic AI matures, the focus has shifted from managing individual agents to authorizing entire “swarms” of autonomous agents that work continuously in the background. This represents a major leap in trust — placing long-running, unsupervised AI systems in charge of real engineering tasks. However, with models improving rapidly, it may be the logical next step to get AI to handle substantial, ongoing work.
Not Entirely New: Recursive Loops in AI
The concept isn’t entirely novel. Recursive loops — where a function calls itself to repeat an action until a condition is met — are a staple of computer science education. AI loops differ by following non-deterministic logic: a sub-agent decides when to stop, rather than relying on a fixed condition. Nonetheless, from the moment programmers started using AI to complete tasks, some form of AI-over-AI recursion was inevitable.
The Simplicity of Agentic Loops
Surprisingly, agentic loops can be maddeningly simple. One popular approach is the Ralph Loop (named after Ralph Wiggum), which summarises all work the model has done and asks if the goal is complete. It’s a straightforward way to prevent AI models from drifting during long tasks — essentially bouncing the model back and forth until the task finishes.
Loops and Test-Time Compute
Loops also align with the broader push for more test-time compute. As OpenAI researcher Noam Brown recently noted, modern models can solve nearly any problem given enough compute. This means one way to guarantee a solution is to keep applying compute until the task is done — especially effective for “hill-climbing” problems like code improvement, where the model makes incremental refinements until it reaches a threshold. In Cherny’s scenario, the loop runs as long as there is compute to spend.
Cost Implications: Token Burn
If that sounds expensive, it should. Like agentic AI before it, AI loops consume tokens far faster than simple Q&A chatbots. And because the loop is designed to run continuously, there is no practical spending ceiling. For companies like Anthropic, which sell tokens, that’s a boon. For everyone else, it may prove a costly way to work.
via TechCrunch AI