AI Coding Agents Raised Commits 180%, Releases 30%: MIT Tracks 100,000 Developers Across Three Tool Generations
A paper posted to SSRN by MIT researchers tracked more than 100,000 GitHub developers across three generations of AI coding tools and found that productivity gains narrow sharply as work moves toward shipping.
The Funnel Numbers
The study measured impact at three stages of the production pipeline:
| Stage | Autocomplete | Interactive Agent | Autonomous Agent |
|---|---|---|---|
| Files edited / commits | +40% | +140% | +180% |
| Projects | — | — | +50% |
| Releases | — | — | +30% |
Across all tool types, AI helped developers create or edit nearly 300% more files. That gain fell to 150% at the code review stage, and to 30% at actual software releases. Marketplace data showed more new apps appearing, but total usage across platforms did not increase — supply rose without a commensurate rise in adoption.
The Weak Links Thesis
The paper frames this as a structural problem. Software production has weak links: the pipeline is constrained by the slowest stages, not the fastest. AI is strongest at accelerating local tasks — writing and editing code files. Human review, coordination, product judgment, testing, and launch processes still determine how much of that output survives to shipping.
The estimated elasticity of substitution is 0.25: for every major improvement in AI coding capability, only a fraction of human work can be displaced, because the work AI handles well is upstream of the decisions and coordination that produce value.
What This Means for Productivity Claims
The finding cuts against the dominant productivity narrative — that more commits equal more output. Companies measuring AI effectiveness by commit velocity or code volume are measuring the wrong thing. The MIT data shows the amplification is real but largely consumed before it reaches users.
The paper was covered by the Financial Times, which noted the same dynamic applies at the team level: AI is raising software supply faster than demand, and much of the additional motion inside companies is being lost before it becomes shipped product.
Context
This is the most methodologically rigorous large-scale study to date on AI coding productivity across the full production funnel. The 100,000-developer sample size and three-generation scope make it more credible than the vendor-sponsored case studies that have dominated coverage. The findings are consistent with earlier work — including a Wharton analysis and a Stanford survey — that found AI productivity benefits concentrated in early-stage individual tasks rather than shipped outcomes.
The elasticity figure (0.25) is the number to watch. If it holds as autonomous agents improve, the implication is that frontier AI capability improvements deliver diminishing real-world productivity returns unless the human coordination bottlenecks are also addressed.