Microsoft Research just published the most rigorous field study of CLI coding agents ever conducted. Tens of thousands of engineers. Four months of real telemetry. One headline number: engineers who adopted Claude Code merged 24% more pull requests. Microsoft also cancelled Claude Code across an entire division the same quarter the paper dropped.
The study — arXiv 2607.01418, authored by Microsoft researchers Emerson Murphy-Hill, Jenna Butler, and Alexandra Savelieva — is the first field study to use direct developer telemetry (not surveys, not controlled labs) to measure both adoption patterns and output impact of agentic command-line tools at enterprise scale. The window covers January through April 2026, when Microsoft made both Claude Code and GitHub Copilot CLI available to its engineers concurrently.
What the Data Actually Shows
The headline metric is a 24% increase in merged pull requests for adopters, and the lift held across the full four-month observation window. This is not a survey asking developers if they felt more productive. It is telemetry of actual code merged into actual repositories.
The secondary finding is more interesting: Claude Code triggered 2.3 times more file edits per session than Copilot CLI. That gap reflects how each tool operates — Claude Code holds a million-token context window and works autonomously across the codebase; Copilot CLI optimizes for speed on discrete shell and GitHub tasks. If you want a tool that touches more of the codebase per session, the data is clear about which one does that.
Before you screenshot that 24% number and paste it into a Slack message to your VP of Engineering, read the next section.
What the Data Cannot Tell You
The researchers used merged pull requests as a proxy for output, and they acknowledge the limitation directly. Merged PRs reward small, frequent commits. A developer who learns to split work into smaller chunks — something AI agents naturally encourage — will show a higher PR count without necessarily delivering more business value.
A separate study tracking 250,000 developers across 60+ enterprises found that AI-generated pull requests wait 4.6x longer in code review and introduce 15–18% more security vulnerabilities. Veracode’s 2026 research puts it more bluntly: 45% of AI-generated code contains OWASP Top 10 vulnerabilities — 2.74x the rate of human-written code. The 24% throughput gain and the 4.6x review backlog are both real. They may cancel each other out. They may not. The Microsoft study is not designed to tell you which.
How Adoption Actually Spreads
The study’s most actionable finding is not the PR count — it is how adoption happened. Top-down mandates were not the driver. Peer networks were.
When at least a quarter of an engineer’s code-review peers had already used the tool, that engineer had 54% higher odds of trying it. When skip-level peers (engineers who share a skip-level manager) had used it, the odds jumped to 216% higher. The signal from visible peer use was the dominant adoption driver, stronger than seniority, team size, or any demographic factor the researchers tested.
If you are rolling out an AI coding tool in your organization, this is the rollout strategy: do not send a company-wide announcement. Identify engineers in dense review networks who are already enthusiastic, give them access first, and let the adoption graph do the work.
Why Microsoft Cancelled It Anyway
Experiences and Devices — the Microsoft division covering Windows, Microsoft 365, Outlook, Teams, and Surface — was instructed to remove Claude Code by June 30, 2026. EVP Rajesh Jha framed the decision as toolchain unification: “We need to shape the tool alongside GitHub — we can’t do that with a third-party product.”
“We need to shape the tool alongside GitHub — we can’t do that with a third-party product.”
Rajesh Jha, EVP, Microsoft Experiences & Devices
The June 30 deadline lands exactly on Microsoft’s fiscal year-end. The cost reduction heading into FY2027 was not incidental. This is not a verdict on Claude Code’s performance. It is a governance and competitive decision — Microsoft owns GitHub, Microsoft sells Copilot, and having tens of thousands of its own engineers on a competing vendor’s CLI is strategically awkward. Claude models remain available through Copilot CLI. The Anthropic partnership, including the November 2025 Foundry agreement, is intact. The Claude Code interface is gone, not the models.
The Broader Problem: Token Bills at Scale
Microsoft is not the only company discovering that AI coding tools scale uncomfortably fast. Uber deployed Claude Code in December 2025. By March 2026, 84% of Uber’s engineers were classified as agentic users — up from 32% in February. The company burned through its entire 2026 AI tools budget by April. Per-engineer costs were running $500 to $2,000 per month. Uber has since capped per-engineer AI tool spending at $1,500 per month, per tool.
This is becoming industry-standard pressure. TechCrunch called it “the token bill comes due” — a moment when enterprises that adopted AI coding tools without spend controls are now scrambling to build them retroactively. Most engineering organizations are three months away from learning this firsthand.
What You Should Do With This
The 24% figure is real data from a rigorous study, and you can use it in internal tool debates. It is the most credible productivity number for CLI coding agents published to date. Cite the paper, not this article.
Do not use it in isolation. Pair the adoption study with quality metrics from your own codebase. Watch review queue depth. Watch security scan results. Watch whether your junior engineers are growing or just shipping AI output they do not understand.
Set spend controls before you deploy, not after you hit the ceiling. And if you are building the case for an AI coding tool rollout, start with the engineers who review each other’s code. The Microsoft data says they will figure it out themselves.
