Engineers Embrace Loop Engineering For AI Agents

Boris Cherny, who leads Claude Code at Anthropic, told CNBC - as reported by Business Insider - that he stopped writing individual prompts: "I don't write the prompt anymore. Claude writes the prompt, and now I'm talking to that new Claude that is kind of coordinating." In a separate public post on X, OpenAI engineer Peter Steinberger wrote: "You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." Writer and engineer Addy Osmani then published "Loop Engineering," a post that named the pattern and described a five-piece architecture including discovery, task decomposition, an orchestration layer, verification, and persistent memory. The New Stack confirmed Cherny's framing: "my job is to write loops." AINews covered the concept under the headline "Loopcraft: The Art of Stacking Loops," describing it as a discussion sparked by Steinberger, Cherny, and Andrej Karpathy in the same week.

Loop engineering, as described across multiple practitioner posts, treats agent interaction as a system design problem. Instead of issuing a single prompt and reviewing the result, the engineer builds a recurring orchestration - a loop - that sets goals, invokes the agent, checks outputs, and re-prompts automatically. Examples cited include a /goal control that keeps an agent working until a task completes. Addy Osmani's five-part architecture maps each component to existing tooling. Multiple authors frame loop stacking - chaining several such loops - as the next layer above individual agent harnesses and scheduling.

Several sources, including Osmani's post, flag token and compute costs as the central operational constraint: loops that run longer and invoke agents repeatedly can burn significantly more compute than a single well-crafted prompt. Observability and verification steps are highlighted as critical to prevent silent failures or runaway cost. Teams adopting the pattern must instrument loop state, set budget guards, and define stopping conditions.

• •Which developer platforms and low-code runtimes ship loop primitives (goal setters, watchdogs, persistent memory) as first-class SDK features.
• •How toolchains add audit trails and cost controls into loop execution.
• •Whether the term 'loop engineering' stabilizes or fragments into competing names (loopcraft, agent orchestration, harness engineering) across communities.