AI-Driven Software Progress Is Already Here, but Unevenly Distributed

Summary

This essay argues that the future of AI-assisted software development has arrived, but its effects are uneven. The author observes that AI is producing code at previously unthinkable speeds and enabling new software, custom scripts, and small projects that were previously impractical. However, familiar large products have not visibly increased their feature velocity at the same pace. The essay cites Harvard research covering about 300 million work events involving 700,000 workers across 700 firms: AI-agent adoption was associated with 30% more lines of code, 20% more commits, and 23% more pull requests, while the increase in features deployed to production was small and statistically insignificant. The author considers deployment delays and a shift toward AI features, but argues that neither explanation accounts for the gap. Drawing on Atul Gawande’s discussion of coordination in complex projects and Fred Brooks’s distinction between accidental and essential complexity, the essay says AI has sharply reduced the work of translating ideas into executable code. It has not eliminated the harder work of deciding what to build, reconciling competing priorities, and coordinating experts and stakeholders. Small teams with a clear vision, or projects with an objective correctness check, can therefore benefit much more than large products with ambiguous requirements. The conclusion is that faster coding does not automatically produce more shipped features; AI may instead allow engineers to spend more time on the essential engineering problems that remain.