AI coding agents tripled commits, but releases rose 30%, study finds

A working paper using data on more than 500,000 GitHub developers finds AI tools raise coding activity sharply, while human steps such as review and release keep shipped software growing far more slowly.

Illustration of small robots pouring a flood of code and papers into a wide funnel while only a few boxes come out the bottom onto a conveyor checked by one developer. Image: AgenticFeed illustration

AI coding agents have sharply raised how much code developers write, but far less of that work turns into shipped software, according to a working paper by economists at MIT Sloan and the Wharton School. Using data on more than 500,000 GitHub developers, the authors estimate that adopting autonomous coding agents on top of earlier AI tools raised weekly commits by a cumulative 240%, while software releases rose about 30%.

The paper, "Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools," is by MIT Sloan's Mert Demirer and Wharton's Leon Musolff, with co-author Liyuan Yang. It circulated as NBER Working Paper 35275 in May and was revised in September. It drew fresh attention on X this week after MIT Sloan published a summary, and NBER notes that its working papers have not been peer reviewed.

Measuring each stage

The authors split software work into stages that run from lines of code and files touched up through commits and pull requests to repositories and releases. They combined public GitHub data with internal Microsoft telemetry to date when each developer adopted AI tools, including GitHub Copilot and OpenAI's Codex, the paper says.

Each adopter was matched with a similarly active developer from the same calendar week one year earlier. The authors say drawing controls from the prior year avoids counting people who were already using AI privately.

The study groups tools by generation. Autocomplete alone raised weekly commits about 30%, according to the paper. Adding sync agents, which work alongside a developer in real time, took the cumulative effect to about 180%, and adding async agents that finish assigned tasks on their own brought it to about 240%.

Gains shrink up the chain

The effects fade at each later stage. For the full set of tools, the abstract says, the 240% gain in commits "falls to 80% for the number of projects, and to 30% for actual releases."

Sync agents raised lines of code by 958% and pull requests by 86%, but releases by only 20%, the paper reports. Autocomplete followed the same pattern, lifting lines of code by 234% but releases by just 9%.

The authors read this as support for what economists call the weak-link hypothesis. AI speeds up some tasks, but output is held back by steps people still do at the old pace, such as reviewing and merging changes. They estimate an elasticity of substitution of 0.23 between AI and human effort, which the paper says "indicates strong complementarities."

More apps, same usage

The researchers checked the result against four major software marketplaces. They found "a sharp increase in the number of new apps but no increase in total usage," the abstract says, and conclude that task-level gains "have therefore translated only partially into shipped and used software thus far."

In the MIT Sloan summary, the authors write that developers still have to test and refine apps until they reach market fit, and that "these higher-level tasks may remain constrained even when AI lowers the cost of writing code." Demirer told MIT Sloan that some bottlenecks "are truly technical and mechanical tasks" that people still have to complete.

He said organizations could respond by making teams smaller and shifting resources toward later stages of development, such as releasing and maintaining products. The paper also finds that the effects were larger for less active developers, though they stayed substantial across the activity distribution.

The MIT Sloan article reports figures from an earlier version of the study, based on more than 100,000 developers, with a cumulative commit effect of up to 180% and 50% more projects. The revised NBER paper uses the larger sample and reports the higher estimates cited above.

Source: NBER, Mert Demirer, Leon Musolff, Liyuan Yang, “Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools”

Additional sources: MIT Sloan, Sara Brown, “AI boosts worker productivity, but does that translate to final outputs?”

Get the AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam, and you can unsubscribe at any time.

Want to reach readers following AI agents? Advertise with AgenticFeed

Next event · 30 Nov, 2026

AWS re:Invent 2026 in Las Vegas

AWS’s annual conference runs Nov. 30 to Dec. 4. Matt Garman’s CEO keynote is on Dec. 1, and the catalog lists more than 2,200 sessions.

The AgenticFeed newsletter

A free, regular roundup of the agent stories that matter. No spam.