# Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI

**Authors:** Emerson Murphy-Hill, Jenna Butler, Alexandra Savelieva  
**arXiv:** [2607.01418](https://arxiv.org/abs/2607.01418)  
**Date:** 2026-07-01  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2607.01418v1/

## Key points

- Study on adoption and impact of command-line tools like Anthropic's Claude Code and GitHub's Copilot CLI conducted by Microsoft researchers
- Usage patterns of tens of thousands of engineers during early 2026 rollout examined
- Tools primarily spread through social networks, retention linked to coding activity rather than demographics
- Adopters merged approximately 24% more pull requests, indicating increased output
- Benefits beyond speeding up typing or code generation: developers took on different types of work, ran parallel streams of activity, more willing to tackle complex tasks
- Identified potential threats to validity: arbitrary thresholds for retention/merge windows, confounds in cross-sectional adoption study, limitations in using merged PRs as proxy for productivity
- Insights into effectiveness of agentic command-line tools in enhancing engineering velocity within organizations
- Importance of considering visible peer use as central component of rollout strategy for such tools

## Summary

A study on the adoption and impact of command-line tools like Anthropic's Claude Code and GitHub's Copilot CLI was conducted by researchers at Microsoft. The usage patterns of tens of thousands of engineers during an early 2026 rollout were examined. The findings showed that these tools were primarily spread through social networks and retention was closely linked to coding activity rather than demographics. Adopters merged approximately 24% more pull requests, indicating a tangible increase in output. The survey results also revealed that using these CLI coding agents had benefits beyond just speeding up typing or code generation. Developers took on different types of work, ran parallel streams of activity, and became more willing to tackle complex tasks they may have previously avoided. However, the study also identified potential threats to validity such as arbitrary thresholds for retention and merge windows, confounds in the cross-sectional adoption study, and limitations in using merged PRs as a proxy for productivity. Despite these challenges, the research provided valuable insights into the effectiveness of agentic command-line tools in enhancing engineering velocity within organizations. It emphasized the importance of considering visible peer use as a central component of rollout strategy for such tools. Although this study was specific to Microsoft in early 2026, its implications may be relevant to other organizations with similar software development environments. It is important to note that while the authors are Microsoft employees with proximity to AI tool development, their research was conducted independently without explicit pressure from their employer.