ResearchNews
NBER: AI coding agents boost commits, not releases
An NBER study of more than 500,000 GitHub developers finds autonomous coding agents raise commits by 240%, but releases rise only 30%. The bottleneck is human.
· 2 min read

What was studied?
An NBER working paper by Mert Demirer, Leon Musolff and Liyuan Yang combines activity data from more than 500,000 GitHub developers with AI usage telemetry (NBER). It compares three tool generations: autocomplete, interactive coding agents and autonomous coding agents. First released in May and revised in September, it is back at the center of developer discussion this week.
The numbers
- Commits: autocomplete +30%, interactive agents +180%, autonomous agents +240% cumulative effect.
- For autonomous agents, the 240% falls to +80% in number of projects and +30% in actual releases.
- Across four major software marketplaces, the number of new apps rose sharply, but total usage did not.
The authors read this as support for the weak-link hypothesis: large task-level gains from AI are absorbed by human bottlenecks further down the production chain. They estimate an elasticity of substitution of 0.23 between AI and human effort, which signals strong complementarity. AI does not replace human work elsewhere in the chain; it depends on it (NBER PDF).
Caveats
This is a working paper, not peer reviewed. Two of the authors disclose that they are paid research consultants for Microsoft, which owns GitHub. Commit counts alone are a weak metric, since an agent can produce many small commits. That is exactly why the paper's push to measure releases and real usage matters.
What it means in practice
If your team works with agents, the speedup probably stalls not in writing code but in review, testing, decisions and the release process. METR finds the length of tasks agents can handle doubles every 3.6 months (METR: AI agent time horizons double every 3.6 months), so this gap will widen. Track release cycle time and review load, not commits or lines written.
Produced with the help of AI, reviewed by an editor. How we make our articles
Reviewed and edited by the Neural Daily desk. About


