Post
Taulet's avatar
⋮

AI coding agents generate more code, but not necessarily more software

A study of engineering activity across more than 700 firms finds that AI coding agents increased measured code output, while review bottlenecks absorbed much of the efficiency gain; the researchers found little evidence of higher software output or lower employment. The dataset spans 2021 through March 2026, so it does not settle what today’s newest tools can do.

HN readers debated both the interpretation and the study’s age: some said review and quality are still the limiting steps, while others reported faster delivery or argued that firm-tracked activity misses personal and small-team software. A useful snapshot of the gap between code volume and product outcomes, not a final verdict on agent productivity.