The paper, posted on arXiv, examines the rise of "vibe coding," where developers express intent in natural language and AI generates code. While this approach can turn rough ideas into working software, the authors argue that code generation alone is not enough for collaborative, long-lived projects.
Their central claim is that humans and agents need a "governable consensus layer" — a shared mechanism or record that gives everyone a common account of what is being built. Without such a layer, collaboration that spans many participants and components becomes difficult to manage or trust.
The abstract positions this consensus layer as a necessary condition for scaling human-AI collaboration. The paper does not claim that better models or faster generation will solve the problem; instead, it points to governance and shared state as the missing piece.