Loop engineering: the era of designing AI's loop, not the prompt

The skill of using AI well is shifting from prompts to loop design. Loop engineering, which rose in mid-2026, is about designing the loop an agent runs on its own, instead of instructing the agent step by step. Andrej Karpathy ran a coding agent for two days, conducted 700 experiments, and found 20 optimizations that sped up training. ASAP summarizes the shift from the primary remarks.

A lineage in three stages: from prompt to loop

The skill of working with AI is now in three stages: prompt, harness, and loop. First was prompt engineering, writing one good prompt. Next was harness engineering, filling the context window with the right information. Now it is loop engineering, designing the loop an agent runs on its own.

Why this shift matters

Line the three stages up and a pattern appears: the point where a human steps in keeps moving later. Prompts needed a hand at every sentence, the harness at every task. A loop needs a person only to set direction and stopping conditions; the machine carries the rest of the repetition. What makes Karpathy's two days of 700 experiments striking is not the count itself but that, across those 700 runs, human effort was spent essentially once, at the moment the loop was designed. The unit of work grew from "one answer" to "one design."

What the "Karpathy Loop" and AutoResearch show

The flagship case of loop engineering is Karpathy's experiment. Andrej Karpathy ran a coding agent nonstop for two days, conducted 700 experiments, and found 20 optimizations that cut training time. His open-source tool AutoResearch implements the same approach directly. You describe research directions in a markdown file and point an agent at the repo, and the agent repeats experiments while keeping only the changes that beat the current best. The human just sets direction and walks away. Results come from the whole loop, not one answer from the model, which is why tools like Codex are becoming agent loop managers rather than chatbots.

Where a practitioner should start

Carrying this over to a working engineering team runs into a snag. Keeping "only the changes that win" assumes a machine can score who won, which requires a clear evaluation function. Tasks that reduce to a number, like cutting training time, hand off to a loop easily; work whose quality resists metrics, like code-review judgment or product-spec completeness, does not. So the first gate for adoption is not a tool but defining what counts as success. That definition is the loop's blueprint.

The trap of loopmaxxing

A good loop needs three things: stopping conditions, verification, and feedback. Without them you fall into loopmaxxing, running a loop endlessly and only inflating the cost. The field, too, is wary of this excess repetition. Turned around, those three are the standard that separates a skilled loop engineer. Someone who builds a loop that runs, verifies, and stops on its own gets more done than someone who asks one good question. The core skill after prompting has moved from instruction to design.

Source: ASAP summary of Andrej Karpathy's "Loopy Era" and AutoResearch remarks (running a coding agent for two days, about 700 experiments, 20 training optimizations found) and 2026 reporting on loop engineering.