Loop engineering is the practice of designing the automated system that prompts an AI agent on your behalf, instead of you typing each instruction by hand. It became a defining idea in AI circles in June 2026, when developer Peter Steinberger posted that people should stop prompting coding agents and start “designing loops that prompt your agents,” a post that crossed 2.2 million views in days.<\/span><\/p>\n Google engineering lead Addy Osmani formalized the concept the next day in an essay that gave it its name and structure. The shift makes sense once you see the progression: prompt engineering is about the words you send, context engineering is about everything the model can see, and loop engineering is about the repeating cycle, act, observe, decide, repeat, that drives an agent toward a goal without a human typing every step.<\/span><\/p>\n It matters most for anyone running AI coding agents or long, multi-step automated tasks. It matters less if you’re doing one-off, single-turn AI requests, where a well-written prompt is still all you need.<\/span><\/p>\n Loop engineering is the discipline of designing the system that runs an AI agent through a repeating act, observe, decide, repeat cycle, instead of a person prompting that agent by hand at every step. If that sounds like a small distinction, consider what happened in the first two weeks of June 2026: a six-word sentence from developer Peter Steinberger, “you shouldn’t be prompting coding agents anymore, you should be designing loops that prompt your agents,” reorganized how a large part of the AI development community talks about working with these tools.<\/span><\/p>\n The post crossed 2.2 million views within days. Reddit threads split into camps, with some developers calling it “the next abstraction layer” and others dismissing it as “a cron job wearing a hat.” Google engineering lead Addy Osmani published a structured essay the next day that gave the idea a name, loop engineering, and a framework: five building blocks plus a persistence layer sitting outside any single conversation.<\/span><\/p>\n This matters for anyone serious about AI right now, whether you’re a student learning how modern AI systems actually get built, or a professional leading a team that’s trying to figure out what skill to invest in next. This guide explains exactly what loop engineering is, why it emerged now, how it connects to prompt engineering and context engineering (concepts you may already know), and what it practically changes about how people work with AI agents in 2026.<\/span><\/p>\n Loop engineering is designing an automated cycle, act, observe, decide, repeat, that drives an AI agent toward a defined goal until a stopping condition is met, instead of a human manually issuing each instruction one at a time.<\/span><\/p>\n Addy Osmani’s own framing, cited widely across the industry in 2026, captures it cleanly: “Loop engineering is replacing yourself as the person who prompts the agent. You design the system that does it instead.” Anthropic’s Boris Cherny, who built Claude Code, put the same idea even more bluntly on stage just two days before Osmani’s essay: “I don’t prompt Claude anymore.”<\/span><\/p>\n This is the distinction that trips up most people new to the concept. A prompt chain, the kind we cover in our companion guide on <\/span>what prompt chaining is<\/span><\/a>, runs a fixed sequence: Step A leads to Step B leads to Step C, in that order, every time.<\/span><\/p>\n A loop is dynamic. According to MindStudio’s 2026 explainer on the topic, “a chain runs steps in a fixed sequence… a loop is dynamic: the agent can revisit steps, adjust based on feedback, or retry with a different approach.” Chains are predictable and easy to trace. Loops are flexible and better suited to tasks where the right path isn’t known ahead of time, like debugging an unfamiliar codebase or exploring an API you’ve never used before.<\/span><\/p>\n Loop engineering didn’t appear out of nowhere in June 2026. Its direct predecessor is a technique called “Ralph,” described by engineer Geoffrey Huntley in early 2026, according to Lushbinary’s 2026 history of the technique. Huntley ran a coding agent inside a plain while-loop: feed the agent the same prompt against a written specification, let it pick exactly one task and implement it, then start a completely fresh agent instance and feed it the identical prompt again.<\/span><\/p>\n He named it after Ralph Wiggum, the perpetually confused Simpsons character, because the technique is, in Huntley’s own words, “deterministically simple in an unpredictable world.” The insight that made it work was the context reset: a long agent session degrades as its working memory fills with old reasoning, dead ends, and stale file contents. Ralph sidesteps that entirely by giving every iteration a clean slate that reads the current state of the project from disk and picks up exactly where the last iteration left off.<\/span><\/p>\n Understanding loop engineering is much easier once you see it as the latest step in a clear progression, not an isolated new idea.<\/span><\/p>\n This is the practice of crafting the words you send to an AI model to get a better output. It’s about communication: what instructions to give, what examples to include, how to format the request. This was the dominant skill from roughly 2022 through 2024.<\/span><\/p>\n In June 2025, AI researcher Andrej Karpathy argued that “prompt engineering” understated the actual work involved in serious AI applications. He proposed context engineering instead, describing it as “the delicate art and science of filling the context window with just the right information for the next step.” Shopify CEO Tobi Lutke echoed the idea, calling it “the art of providing all the context for the task to be plausibly solvable by the LLM.”<\/span><\/p>\n The distinction matters: prompt engineering is about the instruction itself. Context engineering is about everything else the model needs to see, retrieved documents, conversation history, tool definitions, memory, structured around that instruction. We go deeper into this exact comparison in our guide on <\/span>context engineering vs. prompt engineering<\/span><\/a>.<\/span><\/p>\n As agents started operating at greater scale in 2026, a third layer emerged. According to Atlan’s 2026 industry analysis, Ryan Lopopolo of the OpenAI Codex team named the shift precisely: “Agents aren’t hard, the harness is hard.” Harness engineering is about the environment an agent runs inside, the tools it has access to, the permissions it operates under, the feedback loops that tell it whether its actions worked.<\/span><\/p>\n This is where the repeating cycle itself becomes the object of design. According to Tosea.ai’s 2026 complete guide, the breakdown is clean: “Prompt engineering is about the words you send; context engineering is about all the information the model sees; harness engineering is about the environment the agent runs in; loop engineering is about the iterative cycle that drives the agent toward a goal.”<\/span><\/p>\n Crucially, loop engineering doesn’t replace the first three layers. As puppyone’s 2026 analysis puts it, “loop engineering doesn’t replace prompt engineering or context engineering. It wraps them.” Every loop still needs well-crafted prompts and well-managed context at each iteration; it just adds the structure that decides when and how those prompts get triggered.<\/span><\/p>\nKey Highlights of Prompt Engineering to Loop Engineering<\/h2>\n
\n
What Is Loop Engineering, in Plain Terms?<\/h2>\n
A Loop Is Not the Same as a Chain<\/h3>\n
The Origin Story: A Simple While-Loop Named After a Cartoon Character<\/h3>\n
The Four-Layer Progression: Prompt, Context, Harness, Loop<\/h2>\n
Layer 1: Prompt Engineering<\/h3>\n
Layer 2: Context Engineering<\/h3>\n
Layer 3: Harness Engineering<\/h3>\n
Layer 4: Loop Engineering<\/h3>\n