NBER study of 500,000+ GitHub developers: coding agents multiply commits and code, but releases rise only ~20%
Combining public GitHub data on more than 500,000 developers with internal Microsoft telemetry, the paper compares autocomplete, sync agents (Copilot, Claude, Codex) and async agents (Copilot, Codex) against matched controls from a year earlier. Task-level output rises sharply with each generation, but gains shrink up the chain from code to releases. The authors attribute most of the gap to experimentation: much more code is written, rewritten or thrown away, with review and release capacity as the bottleneck.
The numbers (as reported)
| Developers studied | 500,000+ |
|---|---|
| Weekly commits | +30% autocomplete, ~+180% with sync agents, ~+240% with async agents (cumulative) |
| Sync agents: lines of code | +958% |
| Sync agents: pull requests | +86% |
| Sync agents: releases | +20% |
| Lines added vs deleted after sync agents | 7.0x added, 12.2x deleted |
Facts
- Company / subject
- NBER working paper (Demirer, Musolff, Yang)
- Tools
- GitHub Copilot, Claude Code, OpenAI Codex
- Codebase
- —
- Work type
- Research (productivity study)
- Stack
- —
- Published
- 2026-05
- Last checked
- 2026-10-19
Caveats
Observational event study, not an RCT. Uses internal Microsoft telemetry, and NBER's disclosures say Demirer and Musolff previously held postdoc positions at Microsoft and Demirer is a paid research consultant to the company. Revised September 2026.