TLDR: A study on the LLM agent assistant Cursor found that its adoption leads to a significant but transient increase in development velocity (lines added and commits) in the first two months. Concurrently, it causes a persistent increase in static analysis warnings and code complexity, indicating technical debt accumulation. This accumulated technical debt subsequently slows down future development velocity, suggesting a trade-off where initial speed gains come at the cost of long-term code quality and maintainability.
Large language models, or LLMs, are quickly changing how software is built. Tools known as LLM agent assistants are becoming very popular, with many developers reporting huge boosts in productivity. However, there hasn’t been much solid evidence to back up these claims, especially regarding the long-term effects on software quality.
A recent study titled “Speed at the Cost of Quality? The Impact of LLM Agent Assistance on Software Development” by Hao He, Courtney Miller, Shyam Agarwal, Christian Kästner, and Bogdan Vasilescu from Carnegie Mellon University, delves into this crucial area. The researchers aimed to understand the real impact of adopting a widely used LLM agent assistant, Cursor, on both how fast projects develop and the quality of the software produced. You can find the full paper here: Research Paper.
To conduct their study, the team used a sophisticated method called “difference-in-differences.” They compared GitHub projects that started using Cursor at different times with a carefully matched group of similar projects that didn’t use Cursor. They identified 807 GitHub repositories that adopted Cursor between January 2024 and March 2025 by looking for specific configuration files. They then matched these with 1,380 control repositories that had similar characteristics and development patterns before any adoption.
Development Velocity: A Short-Lived Boost
The study found that adopting Cursor led to a significant, but temporary, increase in development velocity. Projects saw a substantial 3-5 times increase in lines of code added during the first month of adoption. This gain, however, quickly faded, largely disappearing after just two months. While there was an initial surge in commits (55.4% in the first month, 14.5% in the second), this also proved to be transient.
This suggests that while LLM agent assistants can provide an immediate burst of productivity, this speed is not sustained over the long term in open-source projects. One possible reason explored by the researchers is an “excitement-frustration-abandonment cycle” where developers initially experiment with the tool for easy tasks but become frustrated when it struggles with more complex scenarios, leading to reduced usage.
Software Quality: A Persistent Decline
In contrast to the fleeting speed gains, Cursor adoption showed a more lasting and concerning trend for software quality. The study observed a significant and persistent increase in static analysis warnings (a measure of potential issues like bugs, maintainability problems, or security flaws) by 30% after adoption. Code complexity also rose significantly by 41%.
This indicates a sustained accumulation of what’s known as “technical debt.” Technical debt refers to the extra work needed in the future due to choosing an easy but suboptimal solution now. The study found that this increase in static analysis warnings and code complexity continued well beyond the initial adoption period, unlike the velocity gains.
The Interplay Between Speed and Quality
Further analysis using advanced statistical models revealed a critical interaction: the accumulation of technical debt, specifically static analysis warnings and code complexity, acts as a major factor causing a slowdown in development velocity in the long run. Essentially, the initial speed boost from Cursor can lead to a more complex and problematic codebase, which then makes future development slower and harder.
Interestingly, the study also found that Cursor adoption inherently leads to a more complex codebase, even when accounting for the overall growth in lines of code. This suggests that the code generated with LLM agents might be structurally more complex than human-written code, creating a “complexity debt” that can amplify frustration and maintenance costs later on.
Also Read:
- RefAgent: A Multi-Agent AI Framework for Smarter Software Refactoring
- Harness Report Reveals AI’s Dual Impact on Software Delivery and Developer Workload in 2025
Implications for the Future of Software Development
These findings have important implications for everyone involved in software development. For developers and teams, the study suggests that simply adopting an LLM agent assistant won’t guarantee sustained productivity. Instead, deliberate process adaptations are needed. This means scaling quality assurance efforts alongside the increased velocity that AI tools can provide. Teams might need to implement more rigorous code reviews, mandate higher test coverage, or use specific “prompt engineering” techniques to guide LLM agents toward higher quality code.
For designers of LLM agent assistants, the research highlights the need to prioritize quality assurance as a core design principle, not just an afterthought. Future tools should ideally suggest tests, flag unnecessary complexities in real-time, and even recommend refactoring. Some radical ideas include tools that automatically reduce their suggestions when project complexity exceeds healthy limits, forcing developers to consolidate existing code before generating more.
While the study provides a valuable snapshot of LLM agent impact in 2024-2025, the technology is rapidly evolving. Future tools and developer practices may address the quality concerns observed. However, this research underscores that truly harnessing AI-assisted software development requires a holistic understanding of the trade-offs between speed, quality, and long-term project health.
