Claude Code Loops are a technique in which Anthropic's Claude Code agent repeats a task in rounds, testing its own output against conditions written in advance and correcting whatever fails, until every condition is met or an attempt limit stops it. For example, an agent building an eCommerce site could write the page in one round, then in later rounds check responsivity, optimize SEO, and page speed.
Geoffrey Huntley, an Australian software developer, described it in July 2025 as one command that hands the same instruction file back to the agent again and again.
%20(1).png)
The technique raises the quality of an output by repetition rather than by better instructions. What the operator writes is a specification: the features the result must have and the standard it must meet, in a form the agent can measure after each round. Boris Cherny, the Anthropic engineer who built Claude Code, said on 2 June 2026 that he no longer prompts Claude and that his job is to write loops.
%20(2).png)
Some critics argue that the technique is an older idea at a larger scale. Research published in 2023 under the names Reflexion and Self-Refine already had a model criticise its own answer and try again. Meghana Harishankara, writing on Medium on 12 June 2026, reports that the most common objection is that an agent amounts to a repeat instruction wrapped around a single model call, an argument that ran past 1,800 comments on Hacker News, a technology forum.
%20(3).png)
