AI Coding Agents Boost Commits by Up to 240%, but Releases Rise Only 30%

News Summary

Two independent research efforts published in 2026 point to the same conclusion: AI coding agents let developers produce far more code, but the extra output is not turning into proportionally more finished, shipped, or widely used software. An Ars Technica report dated October 10, 2026 highlights a Harvard-led study of software firms, and a separate National Bureau of Economic Research (NBER) working paper reaches a strikingly similar picture using GitHub data.

The Harvard study: more code, same outcomes

Researchers Fiona Chen and James Stratton of Harvard University analyzed aggregated data from Jellyfish, a platform that measures engineering team output. The dataset covers roughly 300 million work events, such as commits and pull requests, plus issue-tracking data for more than 700,000 employees at over 700 software firms. The period runs from 2021 through March 2026.

To work out when each firm adopted AI tools, the team combined directly measured AI usage with analysis of GitHub activity. They separated two categories: AI coding assistants, which autocomplete code that humans largely write, and AI coding agents, which write and submit code autonomously from prompts. They then used a difference-in-differences regression, comparing outcomes before and after adoption across firms that adopted at different times.

What changed after firms adopted coding agents

Average firm-level results after adopting coding agents were as follows:

• Total lines of code generated rose by about 30%.
• Total commits rose by about 20%.
• Pull requests rose by about 23%.
• The resolution rate for issues and epics tracked in tools such as Jira showed no statistically significant change.
• There was no detectable shift in the size or complexity of tracked issues, which makes it harder to argue that teams simply took on smaller or easier work.

Where the extra code goes: the review bottleneck

According to the report, the additional output was absorbed downstream. Code review took longer, pull requests were more likely to need revisions, and reviewers left more comments. The authors found little evidence that firms increased overall software output or reduced employment as a result of using these tools. Developers generally know AI-generated code needs careful checking, so the human effort needed to verify it is substantial.

A second view from NBER

An NBER working paper, "Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools" (Working Paper 35275), by Mert Demirer, Leon Musolff, and Liyuan Yang, looks at the same question from a different angle. The paper was issued in May 2026 and revised in September 2026. It uses data on more than 500,000 GitHub developers together with AI usage telemetry, and applies a matched event-study design.

The authors find that each generation of tools raised commits significantly. Cumulative effects were about 30% for autocomplete, 180% for interactive coding agents, and 240% for autonomous coding agents. The gains shrink along the production chain, though. For autonomous agents, the 240% commit effect falls to about 80% for projects and about 30% for releases. Across four major software marketplaces, the number of new apps rises sharply, but total usage does not increase.

Note that an earlier version of this paper was summarized in some secondary coverage with a smaller developer sample and a 180% headline commit figure; the figures above reflect the most recent revision listed on the NBER page.

The weak-link explanation

The NBER authors interpret the pattern through a "weak-link" lens, in which the slowest human-dependent step limits the whole pipeline. They estimate an elasticity of substitution of 0.23 between AI and human effort, which indicates strong complementarity: speeding up one part of the process helps little when review, integration, and release still depend on people. This lines up with the Harvard finding that review time grows when more code arrives.

Wider context and caveats

Other 2026 commentary describes a "productivity paradox", where faster pull requests raise bugs, review time, and maintenance load unless release processes keep up. Some vendor-reported figures circulate, such as companies shipping code 30% faster with agents, but these come from vendor material and we could not verify them independently. A 2025 randomized trial by the research group METR found that experienced developers working on complex, mature codebases were about 19% slower with AI tools, a result often cited alongside these newer studies.

Readers should also keep in mind that both studies measure averages across many teams. Individual teams with strong automated testing, clear specifications, and fast review practices may see different results, and the tools themselves keep changing quickly.

What this means for learners and teams

For students and working developers, the lesson is that writing code is only one stage of building software. Skills in reading code, writing tests, reviewing changes, and designing systems become more valuable as code generation gets cheaper. For engineering teams, the studies suggest investing in review tooling, automated testing, and release pipelines rather than measuring success by lines of code or commit counts alone.