Already 3 months old when the scout caught it.
Loop engineering
What it is
Loop engineering is the practice of designing the surrounding system that runs an AI agent repeatedly and autonomously toward a goal, rather than prompting it by hand one step at a time. Instead of writing a clever single prompt, you build a loop: something finds work, hands it to the agent, checks the result, records progress, and decides what happens next — with the human reviewing outcomes instead of babysitting each step.
Why it's emerging now
Coding agents got reliable enough over the past year that people started leaving them running unattended for long stretches — but doing that well turned out to require real infrastructure (scheduling, isolated worktrees, verification subagents, codified skills) rather than a better prompt. Addy Osmani named and structured this practice in a June 2026 post, synthesizing ideas already floating around Anthropic and other practitioners, and it was picked up fast — O'Reilly Radar republished it within two weeks, and by August it was one of the most-repeated terms at the AI Engineer World's Fair.
How it relates to what you know
It's the natural successor to prompt engineering and context engineering, already on your ledger: those disciplines shape a single call to a model, while loop engineering shapes the repeating cycle of calls around it. It also draws directly on agent harnesses, multi-agent orchestration/subagents, and agent memory — Osmani's five components (automations, worktrees, skills, connectors, subagents) are essentially a checklist for combining those existing pieces into something that runs itself.