A large-scale Microsoft study found that developers who used command-line AI coding agents merged roughly 24% more pull requests, the code changes developers submit for review, than they otherwise would have over a four-month period.
The study, published July 1, found that the productivity boost only appeared when developers actually used the tools regularly. Simply giving people access to the software did not produce the same results.
AI coding agents helped most when developers kept using them
The Microsoft paper covered the company’s early-2026 rollout of two command-line agents: Anthropic’s Claude Code and GitHub Copilot CLI. These tools work in the terminal rather than inside a code editor, helping developers run coding tasks with less step-by-step instruction.
The paper measured a +24.0% increase in merged pull requests per engineer per day. The researchers reported a likely range of +14.5% to +33.7%, and the gain did not fade during the four-month study window.
Microsoft’s researchers also ran a test to check whether the result might have been a fluke. They pretended the rollout had started earlier than it really did and found no similar jump in pull requests. That supported the finding that the real tool use was tied to the increase.
Still, the study does not prove that every AI coding-agent rollout will produce the same gain. The developers who used the tools chose to do so, and the study measured merged pull requests, not software quality, customer impact, security, or long-term maintainability.
The biggest gains came from regular use. Engineers who used the tools five or more days per week saw a lift of more than 50%, compared with roughly 15% for those who used them three days a week.
That detail matters for IT and engineering teams weighing AI coding tools. A license count can make adoption look stronger than it really is. The Microsoft study suggests teams should also watch how often developers actually use the tools and whether use continues after the first trial period.
Copilot CLI users saw about 2.2 times the pull-request lift of Claude Code users during comparable weeks, according to the Microsoft paper. But the researchers said that result came from Microsoft’s own environment and should not be treated as a general ranking of the two tools.
The findings add data to a fast-moving AI development market. OpenAI’s Codex agent can code, fix bugs, and write tests for specific tasks, while newer tools are also pushing AI coding agents beyond the desktop. Cursor’s iPhone app shows how mobile interfaces are becoming part of the same coding-agent wave.
More pull requests can also mean more review pressure
The Microsoft findings landed alongside a separate July 2 enterprise study that tracked 802 developers and 196,212 pull requests at a mid-sized company from January 2024 through April 2026.
At that company, the CTO had publicly announced a formal “2x mandate” in mid-2025, using merged pull requests per engineer per month as the progress metric. By April 2026, per-developer throughput had reached 2.09 times the pre-mandate baseline, rising from 21.2 to 44.3 merged pull requests per active developer.
The gains were not evenly spread. Output grew most in newer repositories, while legacy codebases saw little lift. The difference did not appear to depend on seniority: individual contributors and principal engineers saw similar patterns. The codebase itself was the bigger dividing line.
That is an important warning for organizations with older systems. AI coding agents may work better in cleaner, newer repositories than in complex legacy environments where context, dependencies, and review standards are harder to manage.
At Microsoft, first use spread mainly through peer and manager networks. Engineers were more likely to try the tools when colleagues or direct managers were already using them. Developers who already coded heavily were also more likely to keep using the tools, while lower-activity developers were more likely to stop.
The review process became a major pressure point in the enterprise study. As AI-authored pull requests increased, the share of pull requests receiving at least one human review fell from 89% to 68%. Automated AI review coverage rose from roughly 19% to 84%, and the workload per reviewer roughly doubled.
The added volume did not stop work from shipping. Merge rates stayed mostly flat, and the revert rate declined. But AI-authored pull requests took about 20% longer to merge after the first human review and 22% longer overall after the mandate.
A separate Claude Code pull-request study, published in the ACM Transactions on Software Engineering and Methodology, found a similar need for human cleanup. Across 567 Claude Code pull requests in 157 open-source projects, 83.8% were eventually merged, but only 54.9% went in without additional changes. The remaining 45.1% required human revision, especially for bug fixes, documentation, and project-specific standards.
For technology leaders, the practical risk is not just whether AI coding agents can produce more code. It is whether review systems, security checks, and team processes can keep up with that code.
That concern connects to broader AI-agent governance. A recent TechRepublic story on the enterprise security gap warned that agent permissions, not just agent capabilities, are becoming a new risk surface.
Before scaling AI coding agents, organizations should track four things: adoption by team, how often each developer uses the tools, how much legacy code is in scope, and whether reviewers can handle the extra pull-request volume.
The Microsoft study covered four months, while the enterprise study followed activity through April 2026. Neither study proves that higher pull-request counts lead to better software over the long term. The clearest lesson is narrower but useful: AI coding agents can raise output, but only when teams have the usage habits, codebases, and review capacity to absorb the extra work.
Also read: A Claude Code espionage campaign shows how agentic AI tools can create new governance risks when enterprise access, connectors, and permissions outpace oversight.
Read the full article here
