Peter Steinberger’s June 7 tweet—“You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.”—garnered 2.2 million views and set off a fight over what “loops” even mean. Many couldn’t define it. The clearest voice came from Boris Cherny, creator of Claude Code. On June 2 at Acquired Unplugged, he laid out that a loop is a tiny program you write to prompt the AI, read its output, decide if the task is done, and if not, prompt again. You stop writing prompts by hand; you write the loop that runs them. Cherny says he’s landed 259 PRs in 30 days with loops doing all the prompting. Loops trace back five years, but today’s version differs sharply from the academic “ReAct” while-loop or 2023’s AutoGPT. Geoffrey Huntley’s July 2025 “ralph loop” reset context every iteration and kept things disciplined. In spring 2026, Codex and Claude Code added a /goal command to run ralph loops until validation. The current, fifth stage goes further: loops are the unit of work. They schedule themselves, supervise other loops in parallel, persist state with git-backed crash recovery, and tick off tasks without human attention. It’s cron scheduling plus an AI decision-maker in the body. Critics call it “just cronjobs with a hat.” Cherny admits his loops run on cron, but points out that cron never had a built-in decision engine. A cron job merely fires a fixed script. A loop invokes a model that reads the current state, decides its next move, validates the result and repeats. Stacking loops creates multi-agent orchestration with durable shared state. To start, Claude Code offers a /loop command—Boris’s example: “/loop babysit all my PRs. Auto-fix build issues, and when comments come in, use a worktree agent to fix them.” His five operational tips stress auto mode for permissions, dynamic workflows, /goal or /loop to nudge the agent, running in the cloud and, crucially, self-verification so loops catch their own mistakes.