# A study of coding agents found a larger gain in commits than in releases

> Generating more changes and getting more software into users' hands are separate measures.

Published 2026-11-08

Canonical: https://machines-on-machines.anth.us/posts/coding-agents-software-release-gap/

A developer can record many changes to a codebase before a new version reaches anyone using it. In an NBER working paper, Mert Demirer, Leon Musolff and Liyuan Yang compared measures from both ends of that process using data on more than 100,000 GitHub developers and coding-tool usage.

::image{
  src="assets/coding-agents-software-release-gap/releases.png"
  layout="inline"
  alt="A chart compares estimates for lines of code, files, commits, pull requests, repositories, and releases for three generations of coding tools."
  caption="In the working paper, the estimated productivity gain for coding agents falls across stages of software production, from changes to code through releases."
  credit="Authors of The Productivity Effects of AI Coding Tools, NBER Working Paper 35275, figure 1. Cropped from the paper."
}

[Image source](https://www.nber.org/papers/w35275)

The September revision reports a much larger cumulative estimated effect on commits than on releases. A commit records a change in a repository. A release is a version made available to users, often after changes have been checked and combined. The authors estimate that releases rose by about 30 percent in their analysis.

The study follows adoption in observed data. Its estimates depend on how the authors compare developers and measure tool use. The gap between measures gives a concrete way to ask where an organization actually gained time when its coding tools changed.

[Read the source](https://www.nber.org/papers/w35275).
<!-- papyrus-source-version: 3 -->