As AI coding assistants turn into more and more adept at writing and debugging software program independently, Google Mind co-founder Andrew Ng believes that probably the most vital shift in software program improvement just isn’t about prompting AI, however somewhat about crafting the suggestions loops that allow AI programs to boost themselves constantly.
In an open letter shared on X, Ng referred to as “loop engineering” one of many greatest new concepts in AI-powered software program improvement. The time period has seen some consideration in current weeks, after being mentioned by Boris Cherny, creator of Claude Code, and Peter Steinberger, creator of OpenClaw, who each argued that AI brokers are essentially altering the best way software program is constructed.
“Loop engineering” is a sizzling buzzphrase after mentions of it by Boris Cherny (Claude Code’s creator) and Peter Steinberger (OpenClaw’s creator) went viral on social media. Loops are actually a key a part of how we get AI brokers to iterate at size to construct software program. On this letter, I’d… pic.twitter.com/bhuRw8lrFC
— Andrew Ng (@AndrewYNg) June 30, 2026
The phrase “loop engineering” has turn into a sizzling buzzword, after a dialogue about it on social media by Boris Cherny (creator of Claude Code) and Peter Steinberger (creator of OpenClaw). Loops have turn into a staple of our strategy to getting AI brokers to iterate at size to generate software program.
Ng defined that fashionable AI improvement more and more revolves round three interconnected suggestions loops that allow software program to evolve by means of repeated iterations somewhat than one-time coding periods.
The primary, which he calls the Agentic Coding Loop, permits builders to offer an AI agent with a product specificati“Trendy AI improvement has three interconnected suggestions loops that enable the software program to evolve over a number of iterations, somewhat than a single coding session,” Ng stated.
Additionally learn: Perplexity CEO Aravind Srinivas says Microsoft formed fashionable information work
The primary one, which he calls the Agentic Coding Loop, lets builders feed an AI agent with a product specification and, if obligatory, analysis datasets. Then the AI develops code, exams it, finds bugs and retains bettering the software program till it meets the required necessities.
This has dramatically elevated the quantity of labor AI coding brokers can do on their very own, says Ng. He cited a current instance of constructing a typing app for his daughter the place an AI coding assistant labored autonomously for nearly an hour — testing its personal output utilizing an online browser and bettering the applying time and again — earlier than he needed to step in.
The second stage, the Developer Suggestions Loop, strikes the developer’s position from bug fixing to product selections at a better degree. “As AI programs get higher at testing and debugging their very own code, engineers can spend extra time on interface design, function prioritization and general product course,” Ng stated.
“Builders usually refine product specs after seeing AI-generated implementations, so this iterative collaboration between people and AI is an integral a part of constructing software program,” he stated.
Ng additionally stated it’s troublesome to exchange human judgment as a result of individuals have what he calls a “context benefit.” AI can analyse buyer knowledge, summarise suggestions and help product planning. However builders nonetheless perceive customers, enterprise objectives and working environments in ways in which present AI programs can’t totally replicate.
The final step, the Exterior Suggestions Loop, pushes improvement past engineering groups by leveraging real-world suggestions from early customers, alpha testing, A/B experiments and manufacturing deployments. Ng stated this course of takes for much longer than AI-driven coding, nevertheless it finally helps inform future product selections and guides the subsequent rounds of improvement.
The shifting workflows are additionally altering the position of the software program engineers, says the Google Mind co-founder. Builders are being requested to do extra than simply write code – they have to join product imaginative and prescient to technical execution and incorporate consumer suggestions into future iterations.
First Printed on