Productivity Evidence Shifts from Code Volume to Delivery and Verification

A study across 718 firms associates AI agents with more code and pull requests but no measurable increase in resolved issues and a 49% longer PR-to-merge time; methods and causal interpretation need scrutiny. A small novice study found higher objective workload during AI-assisted code comprehension without measured comprehension gains. AlphaProof Code offers a separate formal-verification example—about 6.5 million agent-written Lean proof lines for scoped libsignal claims—with human-reviewed specifications and undisclosed cost.