Last year, the question was whether AI coding tools actually help. Now we have real data - and the answer is "yes, but you're going to pay for it later."
A new study analyzing the AIDev dataset across GitHub repositories measured what happens to both development speed and code quality when teams adopt AI coding agents (tools like Cursor, Copilot, or Claude Code that can write and modify code autonomously, beyond simple autocomplete). The findings split cleanly into good news and bad news.
Speed Gains Are Real, but They Fade
For teams new to AI tools, coding agents delivered a clear velocity boost. Projects that had never used AI assistance saw meaningful acceleration when agents were introduced.
But here's the catch: teams already using IDE-based AI assistants (like Copilot's inline suggestions) saw a brief spike in activity after adopting full agents, followed by velocity dropping back to near zero - and eventually going negative. The researchers attribute this to higher coordination costs. When an agent generates large volumes of code quickly, someone still has to review, integrate, and maintain all of it. For teams already working with AI, the marginal gain from more autonomous agents didn't stick.
The Quality Problem Is Universal
Regardless of whether a team was new to AI or already experienced, code quality degraded consistently:
- Static analysis warnings increased roughly 18%. These are the automated flags for potential bugs, style violations, and security issues - the kind of problems that compound over time.
- Cognitive complexity rose approximately 39%. Cognitive complexity measures how hard code is to understand and maintain. A 39% increase means code that's significantly harder for humans to work with after the agent writes it.
- Comment density went up in some repos, but not enough to offset the complexity growth.
This matches what many developers report anecdotally: AI-generated code works, but it tends to be verbose, over-engineered, and harder to maintain than what a careful human would write.
What This Means for Your Workflow
The researchers' recommendation is straightforward: pair AI coding agents with strong quality safeguards. That means mandatory code reviews, regular refactoring passes, solid test coverage, and prompt engineering that explicitly prioritizes maintainability over speed.
This isn't an argument against using AI coding tools. The velocity gains for greenfield work are real. But treating agent output as finished code - merging it without careful review - is building a maintenance burden that will slow you down later. The 39% complexity increase doesn't hurt today. It hurts six months from now when you're debugging code that no one on the team fully understands.