On October 9, 2026, Ars Technica published an article analyzing the updated study by Demirer, Musolff and Yang: AI-powered coding agents are sharply increasing how much code developers write, but the number of finished software products is not growing at the same rate.
The updated version of the study was published on September 6, 2026 as an NBER working paper. The authors studied more than 500,000 GitHub developers for whom AI usage telemetry was available — observation at this scale gives the research particular weight.
The study design is a comparative event study. In this approach, developers who adopted AI tools are compared with similar developers who did not, so the observed difference is attributed to the tool itself. Ars Technica relies on the updated version of this working paper and adds new numbers to the debate around coding agents; the primary document is the NBER working paper.
How Three Types of AI Tools Performed
The study measured three types of AI tools separately: autocomplete, interactive coding agents and autonomous coding agents. Their cumulative impact on coding activity — that is, the number of commits — was 30%, 180% and 240% respectively.
The numbers show that the more autonomous the tool, the more coding activity grows. The simplest tool — autocomplete — delivered a 30% increase, interactive agents nearly tripled activity, and autonomous agents more than tripled it. This gap is tied to how deeply the tools are embedded in the developer workflow: the more independent steps agents take, the faster the volume of written code grows.
At the same time, the number of commits is only an activity metric. It measures how much code a developer wrote, but not how much of that code became a working, tested product delivered to users. The study quantified exactly this gap.
From Code to Release: The Effect Fades
The effect weakens at later stages of the chain. The figure for the number of projects is 80%, and for actual releases it falls to 30%. In other words, while the volume of written code jumped sharply, the rate at which it turns into finished products is far lower — the initial 240% surge shrinks to 30% by the release stage.
Autocomplete, interactive coding agents and autonomous coding agents cumulatively raise coding activity (commits) by 30%, 180% and 240% respectively, but the effect drops to 80% for projects and 30% for actual releases.
— SSRN, working paper abstract 6843118 (September 6, 2026)
The study did not analyze the cause of this disconnect separately, but the numbers themselves are unambiguous: writing code has become cheap, while finishing a project — testing, documentation, release preparation — still demands labor. The number of commits grew; the number of releases stayed nearly flat.
In its commentary, Ars Technica presented this situation in the context of the ongoing debate about the effectiveness of coding agents: some call the tools a productivity revolution, while others say the measurement criterion itself was chosen poorly. The updated study added empirical numbers to the debate — now both sides can argue while looking at the same table.
The numbers reveal another aspect: all three tools deliver their largest impact at the earliest stage of the work. The acceleration of code writing already falls nearly threefold at the project level and eightfold at the release level. This shows the gap between the strength of AI tools — text generation — and the complexity of the software delivery process.
AI and Human Labor
The study estimated the elasticity of substitution between AI and human labor at 0.23. Since the elasticity is below 1, this means strong complementarity: AI tools are not displacing human labor but amplifying it. Developers began writing more code with the help of the tools, but final decisions and responsibility stayed with people.
A similar picture was recorded across app marketplaces. In four major app stores, the number of new apps grew, but total usage volume did not increase. New apps are appearing, but their overall consumption is not growing — most new apps never reach a broad audience.
These two findings complement each other: in both programming and app marketplaces, AI tools are increasing production volumes, but growth in final consumption and delivery is much slower. The authors based this conclusion on real telemetry from more than 500,000 developers, which makes the results more reliable than small experiments. An elasticity value of 0.23 is far below 1, confirming that AI and human labor are strong complements — the degree of substitutability is low. So as the tools spread, the role of the human developer is not disappearing; on the contrary, final judgment and quality control are becoming even more important.
