The video discusses the rise of loop engineering in AI, where systems iteratively prompt models to refine outputs until a satisfactory result is achieved, highlighting challenges like biased self-verification and high resource costs. It also notes the emerging shift to graph engineering, cautioning that both approaches currently lack reliable verification and practical scalability, making them expensive experiments rather than proven solutions.
The video explores the emerging practice of loop engineering in AI development, where instead of manually prompting AI models, developers create systems that prompt the AI repeatedly until a perfect result is achieved. This approach evolved from early prompt engineering, which involved crafting precise instructions to get desired outputs from models. As AI models improved, especially with the introduction of agents capable of calling tools and running code autonomously, loop engineering became a way to automate iterative refinement. The practice gained traction after Anthropic’s Claude code creator, Boris Journey, declared he no longer prompted Claude directly but wrote loops to do so, sparking widespread interest and official support from Anthropic.
Loop engineering relies heavily on a verification step, where the AI grades its own output to decide whether to continue iterating. However, this self-assessment often suffers from bias, with models frequently overestimating their improvements. A notable example is Matt Schumer’s “Claude of Duty,” a Call of Duty-style game generated through a gauntlet loop involving a builder agent and a blind critic agent. Despite extensive looping, the critic consistently rated the AI-generated game far below the original, and the loop never reached a satisfactory stopping point. This highlighted the limitations of current loop engineering, where the exit condition’s reliability is crucial but often flawed.
The practice has become resource-intensive, with some developers spending thousands of dollars in tokens to run these loops, often on subscription plans that mask the true cost. Companies like Anthropic have increased rate limits and offered free access to encourage adoption, effectively seeding a consumption habit. Prominent figures in the field, including those at Anthropic and OpenAI, have invested heavily in loop engineering, though skepticism remains about its practical value. Many showcased projects are from-scratch game clones or toy examples, with no proven success in real-world, maintainable codebases involving actual users.
By mid-2024, the hype around loop engineering began to wane, replaced by the concept of graph engineering—a more complex system of wiring AI agents into nodes and edges to manage workflows. This shift was initially a joke but quickly gained serious attention, spawning courses and tools. Unlike loops, graph engineering lacks substantial evidence of effectiveness or widespread adoption. Early benchmarks suggest it may consume more resources for marginal gains. The rapid renaming and rebranding reflect the fast-evolving nature of AI development practices, even as the underlying challenges of verification and cost remain unresolved.
Ultimately, the video cautions viewers to critically assess the viability of loop or graph engineering before investing heavily. The key questions are whether a machine can reliably verify when a task is truly done and who bears the cost of the enormous token consumption. Without reliable verification, these methods risk becoming expensive experiments that fail to outperform skilled human developers. The current state of loop engineering is promising in specific scenarios where automated checks exist, but it has yet to prove itself as a broadly effective replacement for human-guided AI development.