Do AI coding agents improve velocity and quality?

I use AI coding agents. I find them useful. They help me refine requirements and designs, fix bugs, generate code, write tests, and handle other things I’d rather not do myself.

AI is big right now. Just look, for example, at the accepted papers of this year’s International Conference on Software Engineering (ICSE). Most are about AI. And honestly, I get it. It’s new, it’s exciting, and it lets us do things that were infeasible or incredibly laborious before.

But do you know what? I’m also getting a bit tired of having to read, hear, and write about AI in software engineering all the time. So in the next few months I’ll be making a concerted effort to find software engineering papers that are not primarily about AI.

For now though, here’s another summary about a short paper that analyses the AIDev dataset to understand the real-world effects of using LLM-based coding agents on software projects.

AI-powered coding tools are rapidly taking over software development. Two distinct paradigms have emerged so far. The first is pre-agentic coding assistants, which integrate into developer environments like Visual Studio Code to provide real-time code suggestions and inline assistance on demand via chats or autocompletion. The second is IDE and web-based coding agents, which are autonomous AI systems that operate at the repository level to generate entire pull requests and implement features with minimal human intervention.

Coding agents can contribute changes at tremendous speed, which naturally raises questions about code quality, maintainability, and technical debt.

In this study, the researchers analysed the AIDev dataset to understand how coding agent adoption affects development velocity and software quality in public GitHub repositories, and whether those effects differ between greenfield and brownfield projects (side note: Technically, it is about whether the effects of coding agent adoption differ between repositories with and without prior IDE-based coding assistants.).

The results show that coding agents substantially speed up development in “agent-first” repositories. For “IDE-first” repositories, the initial effect is minimal: there’s a short bump of additional activity around adoption, but velocity soon falls back to near zero and eventually goes negative. This suggests that agentic tools work well for workflows new to AI, but yield diminishing returns in AI-saturated ones – possibly because of higher coordination and integration costs, and because greater codebase maturity limits how aggressively agentic changes can be merged.

When it comes to software quality, no such difference exists. Agentic tool adoption is consistently associated with increased maintainability risks regardless of project type. Static analysis warnings rise by about 18%, and cognitive complexity by roughly 39%.

Given the weak and sometimes negative velocity gains, this suggests that agents accelerate the introduction of code that raises long-term cognitive and maintenance load. Duplication effects (side note: Due to code that is “copy-pasted” throughout a project) are small and inconsistent, suggesting that quality risks stem mainly from structural complexity. With regard to comment density, agent-first repositories show minimal effects, while IDE-first repositories show substantial and sustained increases — but not enough to offset the continuous growth in complexity.

Overall, these findings suggest that velocity gains from AI adoption must be paired with strong quality safeguards to prevent the accumulation of technical debt through regular reviews, routine refactoring, comprehensive testing, and incorporating maintainability metrics directly into agent planning and prompting.