Google Brain co-founder Andrew Ng says AI won't replace developers, explains 'loop engineering'

As AI coding assistants become increasingly adept at writing and debugging software independently, Google Brain co-founder Andrew Ng believes that the most significant shift in software development is not about prompting AI, but rather about crafting the feedback loops that enable AI systems to enhance themselves continuously.In an open letter shared on X, Ng called “loop engineering” one of the biggest new ideas in AI-powered software development. The term has seen some attention in recent weeks, after being discussed by Boris Cherny, creator of Claude Code, and Peter Steinberger, creator of OpenClaw, who both argued that AI agents are fundamentally changing the way software is built.“Loop engineering” is a hot 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 now a key part of how we get AI agents to iterate at length to build software. In this letter, I’d… pic.twitter.com/bhuRw8lrFC— Andrew Ng (@AndrewYNg) June 30, 2026 The phrase “loop engineering” has become a hot buzzword, after a discussion about it on social media by Boris Cherny (creator of Claude Code) and Peter Steinberger (creator of OpenClaw). Loops have become a staple of our approach to getting AI agents to iterate at length to generate software.Ng explained that modern AI development increasingly revolves around three interconnected feedback loops that enable software to evolve through repeated iterations rather than one-time coding sessions.The first, which he calls the Agentic Coding Loop, allows developers to provide an AI agent with a product specificati“Modern AI development has three interconnected feedback loops that allow the software to evolve over multiple iterations, rather than a single coding session,” Ng said.Also read: Perplexity CEO Aravind Srinivas says Microsoft shaped modern knowledge workThe first one, which he calls the Agentic Coding Loop, lets developers feed an AI agent with a product specification and, if necessary, evaluation datasets. Then the AI develops code, tests it, finds bugs and keeps improving the software until it meets the specified requirements.This has dramatically increased the amount of work AI coding agents can do on their own, says Ng. He cited a recent example of building a typing app for his daughter where an AI coding assistant worked autonomously for almost an hour -- testing its own output using a web browser and improving the application over and over -- before he had to step in.The second stage, the Developer Feedback Loop, moves the developer’s role from bug fixing to product decisions at a higher level. “As AI systems get better at testing and debugging their own code, engineers can spend more time on interface design, feature prioritization and overall product direction,” Ng said.“Developers often refine product specifications after seeing AI-generated implementations, so this iterative collaboration between humans and AI is an integral part of building software,” he said.Ng also said it is difficult to replace human judgment because people have what he calls a “context advantage.” AI can analyse customer data, summarise feedback and support product planning. But developers still understand users, business goals and operating environments in ways that current AI systems cannot fully replicate.The last step, the External Feedback Loop, pushes development beyond engineering teams by leveraging real-world feedback from early users, alpha testing, A/B experiments and production deployments. Ng said this process takes much longer than AI-driven coding, but it ultimately helps inform future product decisions and guides the next rounds of development.The shifting workflows are also changing the role of the software engineers, says the Google Brain co-founder. Developers are being asked to do more than just write code – they must connect product vision to technical execution and incorporate user feedback into future iterations.