Andrew Ng, cofounder of Google Brain, published an open letter laying out a three-loop framework for building software products with AI agents. The letter responds to the viral spread of “loop engineering” as a concept, a term that gained traction after Claude Code creator Boris Cherny and OpenClaw creator Peter Steinberger suggested traditional coding is giving way to agent-driven development, Times of India reported.
Ng identifies three distinct loops, each operating at a different speed.
The Three Loops
The agentic coding loop is the fastest. An AI agent receives a product specification and an optional set of evaluation tests, writes code, tests its own output, finds bugs, and iterates until the software meets the spec. Ng wrote that this loop executes every few minutes, with the agent building and testing new versions of the software without human intervention. He described using a coding agent over a weekend to build a typing-practice app for his daughter, with the agent working for roughly an hour at a stretch, checking its own output in a web browser multiple times before reporting back.
The developer feedback loop operates on a timescale of tens of minutes to hours. The developer reviews the current product and steers the agent toward improvements. Ng noted that this role has shifted over the past year: developers used to spend most of their time acting as QA for coding agents, manually finding bugs and asking the agent to fix them. With agents now better at testing their own code, developers can focus on higher-level decisions about features, UI design, and user flow.
The external feedback loop is the slowest, involving real users. This includes alpha testing, A/B testing in production, or simply asking friends to try the product. This data informs the developer’s product vision, which feeds back into the specifications given to the coding agent.
The Context Advantage Argument
Ng’s core claim is that humans retain what he calls a “context advantage” over AI systems. Developers know more about users, operating constraints, and product context than the AI does. He avoids the common framing of this human contribution as “taste,” preferring the context-advantage framing because it points toward a concrete path for improvement: help AI systems acquire more context, and the human role shifts accordingly.
“So long as the human knows something the AI does not, human-in-the-loop is needed to inject that knowledge into the system,” Ng wrote, according to Times of India.
This positions loop engineering not as the death of coding but as a restructuring of what developers do. The inner loop (writing and debugging code) is increasingly automated. The outer loops (product direction, user validation) remain human-driven. For teams building agent-based products, the implication is that optimizing each loop independently matters more than optimizing the model itself. The harness, the evaluation set, and the feedback pipeline determine how effectively the model’s output translates into a working product.