Autonomous coding tools were associated with a 240% cumulative increase in commits in a large developer study, but the estimated gain fell to 80% for projects and 30% for releases. The gap shows why a busier repository is an incomplete measure of software productivity.[1]
Mert Demirer, Leon Musolff and Liyuan Yang combined GitHub activity from more than 500,000 developers with AI-use telemetry. Their NBER working paper, issued in May and revised in September 2026, uses a matched event-study design to compare adoption of successive generations of tools. The reported cumulative commit effects were 30% for autocomplete, 180% for interactive agents and 240% for autonomous agents.[1]
Three meanings of productivity
A commit records a change in version control. A project measures activity at a broader level, while a release marks a shipping event. Each has limitations: the same work can be split into several commits, and a release can be trivial or substantial. Reading the outcomes together reveals where a generation tool’s immediate effect becomes a smaller change in delivered software.[1]
The authors interpret the pattern through a production chain in which AI and human effort complement each other. Faster initial coding leaves design review, testing, integration and coordination to complete. A queue at those stages can limit the final output even when the first stage becomes much quicker.[1]
Activity gains diminish downstream
| Outcome measured | Reported cumulative increase | Interpretation |
|---|---|---|
| Commits with autonomous agents | 240% | Earlier-stage coding activity |
| Projects | 80% | Broader software-development output |
| Releases | 30% | Shipped software events |
What the comparison can establish
This was an observational study rather than random assignment. Matching and the event-study design improve the comparison, but adoption can still coincide with changes in projects, motivation or working conditions. The release measure also leaves out forms of value such as better reliability or easier maintenance. The estimates describe the study’s outcomes, not a universal multiplier for engineering teams.[1]
For a team using coding agents, the corresponding question is where draft changes wait. Lead time through review, test failures, rework and rollbacks can show whether additional generated code is helping delivery or adding to an existing queue. Those measures need to sit alongside commit counts.[1]
The paper also examines four major software marketplaces, reporting more new applications without an increase in total usage. That adds a demand-side question to the production story: shipping more software does not ensure that people use more of it.[1]
The practical implication is to evaluate the whole route from a proposed change to software that users rely on. A coding assistant can create valuable prototypes and reduce effort even when releases rise modestly. The evidence asks organisations to measure those benefits where they occur, while investing in the testing and coordination that let more code become a useful product.[1]
Sources & further reading
- Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools National Bureau of Economic Research · Accessed 26 Oct 2026