A way of running AI coding assistants lasted about six weeks in the summer of 2026. Engineers named it in early June and had declared it finished by mid-July.
It started with Boris Cherny, the engineer who built Claude Code, Anthropic’s AI coding assistant. For about two years, using a coding assistant meant typing each instruction by hand and reading each result. In early June, Cherny said he had stepped back from that habit and no longer typed instructions into the assistant. Instead, he wrote small systems to sit between him and the assistant. Each one sends new instructions on its own, checks the results and tries again until a task is finished. He called them loops, and writing them, he said, had become his job.
Addy Osmani wrote about the shift and gave it a name, loop engineering. For weeks, the name held. Guides spread across social platforms and company newsletters, and engineers rebuilt their own versions of Cherny’s setup over weekends.
Then, on July 18, the engineer Peter Steinberger posted a mostly joking question on X. “Are we still talking loops or did we shift to graphs yet?” he asked. A few hours later, Hamel Husain, another engineer, answered with an article titled “Loop Engineering Is Dead. Enter Graph Engineering.” By the end of the weekend, the same online spaces had filled with diagrams, guides and courses explaining why loops were already outdated. Weeks earlier, they had filled with almost identical material explaining why loops were the future.
The speed comes from how little it costs to test an AI engineering claim. An engineer who hears about a new looping method on a Friday can build a version by Saturday night. By Sunday, the loop never stops, loses track of its own progress or drifts from the original task. Thousands of engineers run something close to the same experiment within days. A shared weakness shows up fast and in public, so the whole field notices it together.
Excitement alone used to keep a new idea going for a while. People would call something promising and plan a test, then often never run it. With loops, the public test finished before the public conversation settled.
Cherny’s original practice, a system instructing itself toward a goal, kept running under the new name the whole time.
