Addy Osmani popularized the term in June 2026, building on Peter Steinberger’s line that “you should be designing loops that prompt your agents” and Boris Cherny’s remark that his job is now to write loops. Where prompt engineering optimizes a single instruction, loop engineering designs the system around the agent: what discovers the work, what condition ends it, what verifies the result, and what state carries over to the next run.
In practice this means combining primitives like goal conditions checked by a separate evaluator, scheduled or interval-based reruns, isolated worktrees for parallel agents, and a maker/checker split between the agent that writes and the agent that reviews. The recurring warning is that a loop amplifies its designer’s intent, so the same machinery produces opposite results in the hands of someone mastering a domain versus someone avoiding it.
