Reproduce “Cheap Code, Costly Judgment: A Case Study on Governable Agentic Software Engineering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
James C. Davis, Paschal C. Amusuo, Tanmay Singla, Berk Çakar, Kirsten A. Davis
A 12-week case study on AI-mediated software development introduces "governance conversion," a process where human judgment transforms agent-produced failures into machine-actionable governance mechanisms embedded in the engineering environment. This approach demonstrates how sustained high-velocity agentic implementation can yield governable software progress without continuous human code inspection.
Key insights
- Sustainable agentic velocity relies on relocating governance from human review into the engineering environment, making it machine-actionable to effectively shape agent work, with deterministic controls capable of compounding velocity.
- As implementation becomes abundant, human work shifts towards problem framing, abstraction discovery, architectural judgment, and governance design, rather than low-level coding.
- The research reveals a paradox where governable agentic velocity is achieved by systematically reducing opportunities for unconstrained agent reasoning through bounded task briefs, deterministic guardrails, and controlled output admission.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.