Claude Code Lead: Agentic Loops 'As Big' a Shift as Agents

Agentic loops are the next paradigm shift in AI coding — and they're already operational inside Anthropic's own engineering org. Speaking at Meta's @Scale conference on June 20, 2026, Claude Code lead Boris Cherny declared the transition from agent-prompted code to agent-prompting-agents code "as big as the step from source code to agents was." This isn't a demo — it's how Cherny's team works today.

What happened

Cherny's answer to whether agentic loops are a hype cycle left no ambiguity: "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. As big as the step from source code to agents was, loops are just as important and as big a step" (TechCrunch(opens in new tab), 2026).

He then described the specific loops running in his own work. Agents that continuously scan for ways to improve code architecture — submitting pull requests like any other engineer, then re-running as the code changes. Agents that hunt for duplicated abstractions across the codebase and unify them. The key detail: these agents never stop running. They're persistent background processes, not one-shot tasks.

The most popular implementation pattern — the Ralph Loop, named for Ralph Wiggum from The Simpsons — works by having an agent summarize everything it's done, check whether the goal is met, and bounce back and forth until the task is complete. As TechCrunch notes, OpenAI researcher Noam Brown observed earlier this month that contemporary models can solve nearly any problem if you throw enough compute at them. Agentic loops are the engineering pattern that operationalizes that observation: keep spending tokens on incremental improvement until a threshold is met — or simply keep spending forever.

Why it matters

Cherny is not speculating. He's describing the engineering org that ships Claude Code — a product used by developers who regularly execute multi-hour autonomous coding sessions. When the person building the tool says agentic loops are the next paradigm shift, that's not marketing. It's the internal operating model.

The signal for the dev-tool ecosystem is unambiguous. Every coding tool — Cursor, Claude Code, Windsurf, Codex — is converging on persistent, autonomous agents that keep improving code long after the developer has moved on. The tools that solve the cost-and-trust problem for continuous background agents will own the next 12 months.

The economics are the wildcard. Loops burn through tokens continuously. Unlike a single Q&A call, an agent that runs 24/7 on a codebase has no ceiling on spend (TechCrunch(opens in new tab), 2026). Anthropic, as a token-selling business, benefits directly — but for every other organization, always-running AI means GPU budget planning most teams haven't started. CI/CD pipelines will need to accept AI-authored PRs. Review workflows must handle code you didn't write — and didn't ask for.

What changes for you

If you're using an AI coding tool regularly, agentic loops are coming whether you opt in or not. Three things to do now:

1. Audit your token burn. If you haven't modeled what continuous background agents would cost you, start. Agentic loops remove the ceiling on AI coding spend.
2. Build review muscle for AI-authored PRs. The code quality bar for agent-generated changes differs from hand-written ones. Your review process needs to account for code nobody asked for — but might still be good.
3. Watch CI/CD integration patterns. The first tools to ship agentic-loop-aware pipelines — approving, merging, and measuring AI-authored PRs at scale — will set the standard for the category.

FAQ

What is an agentic loop? An agentic loop is a pattern where an AI agent repeatedly evaluates its own output, decides whether the goal is met, and continues iterating — either until completion or indefinitely. Unlike traditional software loops with explicit stop conditions, agentic loops use the AI's own judgment to determine when to stop, making them suitable for open-ended improvement tasks like code refactoring.

How is this different from what coding agents already do? Most coding agents today operate in one-shot or short-session mode: you give a prompt, the agent generates code, the session ends. Agentic loops are persistent background processes — they run continuously, submitting improvements as the codebase evolves, without being re-prompted.

Which tools support agentic loops today? Claude Code is the most mature implementation, with Cherny describing internal teams running 24/7 loop agents. Cursor's agent mode handles multi-step autonomous sessions within the editor, and both Windsurf and Codex are building toward similar persistent-agent architectures. The category is still emerging — expect rapid feature launches across all major tools this quarter.