Boris Cherny, the creator of Claude Code, used a Friday appearance at Meta's @Scale conference to argue that agentic 'loops' — swarms of AI agents prompting other agents to write code, running continuously in the background — are the next structural shift in how software gets built. Asked from the audience whether loops were a hype cycle or for real, Cherny answered: 'Yes, they're for real.' He placed the shift on a two-year arc that has moved from hand-written source code, to agents writing code, to agents now prompting agents that write the code.
Cherny's framing puts loops on the same level as the jump from human coding to agentic AI. That is a strong claim from someone who built one of the most-used agentic coding products on the market, and it lands at a moment when most developers are still working out how to manage a single agent without it drifting off-task.
Around the 32-minute mark of his talk, Cherny got specific about what he actually runs. One agent continually scans his code for architectural improvements. Another looks for duplicated abstractions that can be unified. Both submit pull requests like any other coder, and because the codebase keeps changing, the agents never stop. There is no end-of-task; the loop is the workflow.
Key facts
- 01Boris Cherny, creator of [Claude](/claude) Code, used Meta's @Scale conference Friday to argue agentic 'loops' are the next step beyond single-agent coding.
- 02Cherny runs continuous loops in his own work: one agent hunts code-architecture improvements, another unifies duplicated abstractions, both submitting pull requests indefinitely.
- 03The shift is framed as a two-year arc — from hand-written source code, to agents writing code, to agents now prompting other agents.
- 04[OpenAI](/openai) researcher Noam Brown said earlier this month that contemporary models can solve nearly any problem given enough compute, the economic logic behind always-on loops.
- 05Token consumption is the catch: loops have no natural stopping point, which suits Anthropic's token-selling business model more than it suits buyers.
The mechanic is older than the marketing. Recursive loops — functions that call themselves until a stopping condition is met — are a staple of introductory computer science. What is new is that the stopping condition is itself non-deterministic: a subagent decides when the work is done, rather than a clean Boolean test. Once developers started letting AI complete tasks, AI supervising AI was an obvious next step.
Some loop patterns are almost comically simple. The Ralph Loop — named after Ralph Wiggum — 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. It is a workaround for the well-documented tendency of agents to lose the plot during long runs. Crude, but it works often enough to be a recurring pattern.
Loops are also the logical conclusion of the industry's push toward more test-time compute. If the model can grind on a problem indefinitely, then for any task that admits incremental progress — code quality, refactor coverage, test suites — the budget, not the model, is the binding constraint. Throw more tokens, get more output. That is the bet.
“contemporary models can solve nearly any problem if you throw enough compute at them”— Noam Brown, OpenAI researcher
Earlier this month, OpenAI researcher Noam Brown put the underlying logic plainly, saying 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 just a way of pre-committing to spend until the problem is solved or the wallet is empty.
The economics cut sharply in Anthropic's favor and against almost everyone else's. Agentic AI already burns tokens faster than chatbot Q&A, and loops by design have no upper bound on token spend — the whole point is that they keep running. For a company in the token-selling business, this is a structurally attractive product shape. For buyers, it is a budgeting problem dressed up as a productivity gain, and it will demand new tooling around spend caps, drift detection, and human review checkpoints before most enterprises can adopt it at scale.
It is also worth being clear about what loops are not. They 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, silent regressions. The Ralph Loop exists because agents get lost. More loops do not fix that — they just amortize the cost across more attempts.
Still, if Cherny is right about the scale of the shift, the AI coding market is about to bifurcate. On one side, point-and-shoot agents that complete discrete tasks under human supervision. On the other, always-on loops that treat a codebase as a living surface to be continuously polished by a swarm of AI workers. The first is how most teams use Claude Code and competitors today. The second is what Anthropic appears to be building toward, and what its pricing model is best positioned to monetize. Whether buyers go along depends less on whether the loops work and more on whether the unit economics survive a CFO review.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
