Anthropic is positioning continuous AI agent loops, where multiple AI agents prompt each other to improve code without stopping, as the next major shift in how software gets built. Boris Cherny, creator of Claude Code, argued at Meta's @Scale conference that this represents a structural leap comparable to the move from hand-written code to agentic AI itself.

What Exactly Are Agentic Loops?

Agentic loops are workflows where multiple AI agents run continuously in the background, each assigned to specific tasks. Rather than completing a task and stopping, these agents keep working indefinitely, submitting pull requests and refining code as the codebase evolves. Cherny described running two agents in his own work: one that continuously scans code for architectural improvements, and another that hunts for duplicated abstractions that can be unified.

The concept builds on recursive loops, a staple of computer science where functions call themselves until a stopping condition is met. What is new is that the stopping condition itself is non-deterministic; an AI subagent decides when work is complete, rather than a clean yes-or-no test. One simple pattern, called the Ralph Loop, summarizes everything the model has done so far and asks whether it has accomplished its goal, bouncing the model back into the task if not.

Why Is This Happening Now?

The push toward agentic loops reflects a broader industry shift toward test-time compute, the idea that throwing more computing power at a problem during inference, rather than during training, can solve harder tasks. OpenAI researcher Noam Brown stated earlier this month that contemporary models can solve nearly any problem if given enough compute. Loops operationalize that observation: if compute is the limiting factor, then a continuously running agent is simply a way of pre-committing to spend tokens until the problem is solved or the budget runs out.

"Contemporary models can solve nearly any problem if you throw enough compute at them," said Noam Brown.

Noam Brown, Researcher at OpenAI

This economic logic favors Anthropic structurally. Agentic AI already burns tokens faster than traditional chatbot Q&A, and loops by design have no upper bound on token consumption. For a company in the token-selling business, this is an attractive product shape. For buyers, it becomes a budgeting problem disguised as a productivity gain.

How to Manage Agentic Loops in Enterprise Settings

• Spend Caps: Set hard limits on token consumption to prevent runaway costs, since loops have no natural stopping point and will continue indefinitely if unchecked.
• Drift Detection: Implement monitoring systems to catch when agents lose focus or begin making unintended changes to code architecture or functionality.
• Human Review Checkpoints: Require human approval before agents submit pull requests or make significant refactors, preventing silent regressions and hallucinated edits.

What Are the Real Limitations?

Loops are not a new model architecture, not a benchmark result, and not a capability jump in any single inference call. They are a workflow pattern built on top of existing agents, and they inherit every weakness of those agents: hallucinated edits, runaway refactors, and silent regressions. The Ralph Loop exists precisely because agents get lost during long runs. More loops do not fix that fundamental problem; they simply amortize the cost across more attempts.

Most developers are still working out how to manage a single agent without it drifting off-task, making the jump to swarms of agents prompting agents a significant leap in complexity. The shift will demand new tooling around spend caps, drift detection, and human review checkpoints before most enterprises can adopt loops at scale.

How Will This Reshape the AI Coding Market?

If Cherny is correct about the scale of this shift, the AI coding market is about to split into two camps. On one side, point-and-shoot agents that complete discrete tasks under human supervision, which is how most teams use Claude Code and competitors today. On the other side, always-on loops that treat a codebase as a living surface to be continuously polished by a swarm of AI workers.

Anthropic appears to be building toward the second model, and its pricing structure is best positioned to monetize it. Whether buyers actually adopt loops depends less on whether the loops work and more on whether the unit economics survive a CFO review. For enterprises accustomed to fixed software costs, the idea of open-ended token consumption may prove to be the real barrier to adoption.