Berkeley RDI proposes a three-level framework for autonomous software development, from code ownership to demand ownership

当编码不再是瓶颈：Berkeley RDI 提出软件自主开发三级框架

A position paper from Berkeley RDI, co-authored by Hao Wang and Dawn Song among others, defines three levels of AI-driven software autonomy: Level I (Code Autonomy) where AI delivers complete pull requests but humans still decide what to build and gate deployment; Level II (Pipeline Autonomy) where AI runs design through deployment and humans only state high-level demands; Level III (Demand Autonomy) where AI decides what to build based on telemetry and system state. The authors note that 16 parallel Claude agents built a working C compiler for under $20K, yet frontier agents still degrade sharply on benchmarks requiring continuous software evolution—they add features but struggle to preserve correctness and architectural coherence. The central argument: autonomy is not a single capability, and the real challenge is preserving human intent as direct control recedes. The post does not provide a deployment timeline but expects high-assurance domains to stay at Level I for a long time while internal tools may approach Level II sooner.

Why it matters: This Berkeley RDI position paper isn't a product launch, but its three-level framework pushes the AI software development conversation from 'can it write code' to 'who decides what to build' — a real cognitive increment for practitioners. The cross-institutional author list (C...