AI speeds up coding far more than it ships finished software
Reviewed by Helen Jones
MIT Sloan published a 2 September 2026 summary of research by Mert Demirer, Leon Musolff, and Liyuan Yang on more than 100,000 GitHub developers. Autocomplete raised coding activity by 40 percent. Adding sync agents took the cumulative activity effect to 140 percent, and async agents to 180 percent. Those activity gains showed up as about 50 percent more projects and 30 percent more releases than developers without AI. After agentic coding tools spread in early 2025, app-store listings rose, but downloads and reviews did not keep pace. The researchers say the extra work shrinks at later human-gated stages such as review, merge, and release. Async agents were not given a release-impact figure because they cannot ship software themselves. The study covers GitHub and app-store outcomes, not general office work.
For years, faster generation was treated as proof that the job was done. Teams counted chats started, drafts produced, and tickets touched, then assumed quality would follow if people just prompted harder. This work shows the opposite pattern. Activity can jump while finished, used output barely moves, because the same people still have to check, approve, and release. The scarce resource is not more AI writing. It is a review step that can keep up with the flood.
Analysis
Treat this as a trap to avoid, not a reason to add another agent. Before you expand generation, write the human-in-the-loop rule for one recurring task: what must be checked, who signs off, and what never ships from a blank chat.
Source note
Pulse published by Collab365 Spaces, reviewed by Helen Jones on . Cite as "AI speeds up coding far more than it ships finished software", Collab365 Spaces. 2 sources referenced.