How AI Products Iterate from Minutes to Days Using Three Loops

Andrew Ng’s latest article on The Batch introduces “loop engineering,” a framework for developing AI agents that structures product creation into three iterative loops: the Agentic coding loop (minutes), the Developer feedback loop (hours), and the External feedback loop (days). The article includes a diagram illustrating the three loops.

Agentic coding loop: Given a product specification and optionally an evaluation dataset, an AI agent writes code, tests its own output, and iterates until the code is bug‑free and meets the spec. The author cites a personal example where a coding agent built a typing‑practice app for his daughter, operating for about an hour, repeatedly checking results in a web browser and returning for further instructions without any human intervention. The loop runs quickly—often every few minutes a new version is built and tested—making it a fast‑moving area of innovation.

Developer feedback loop: In this stage the developer reviews the current product and guides the coding agent toward improvements. Previously developers acted as QA, manually finding bugs for the agent to fix. As agents become capable of self‑testing, the time spent on QA drops, freeing developers to make higher‑level decisions such as which core features to add, UI refinements, and user‑flow adjustments. The loop typically spans tens of minutes to several hours. The author describes adjusting the visual design, unlockable cat outfits, and adult login flow in the same typing‑practice app based on developer feedback.

External feedback loop: This slower loop involves broader strategies such as soliciting feedback from friends, releasing to alpha testers, or conducting A/B tests in production. These activities can take days to weeks. The collected data influences the developer’s vision, which in turn refines product specifications that drive the coding agent. As coding agents accelerate development, engineers increasingly assume product‑management responsibilities, balancing vision shaping with user‑feedback integration. The author notes that the most challenging aspect for engineers transitioning to this role is maintaining a clear product vision while iterating between building and gathering feedback.

The article concludes with a promise to publish further guidance on implementing these loops and encourages engineers to embrace broader roles, similar to how product managers and designers are now engaging more directly with engineering work.

References

Strategy: https://www.deeplearning.ai/the-batch/how-to-get-user-feedback-to-your-ai-products-fast?utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&_hsenc=p2ANqtz-9IU5WWvlZnFv4GFOY2DPAg8PWrnhPB_aFANXt2LIg2tC97Y7bMJ07wp5JEddrBEnV27XVg

Developer vision: https://www.deeplearning.ai/the-batch/how-to-get-through-the-product-management-bottleneck?utm_campaign=The%20Batch&utm_source=hs_email&utm_medium=email&_hsenc=p2ANqtz-9IU5WWvlZnFv4GFOY2DPAg8PWrnhPB_aFANXt2LIg2tC97Y7bMJ07wp5JEddrBEnV27XVg

Original article: https://www.deeplearning.ai/the-batch/issue-359