A SaaS company with a mature Ruby on Rails codebase and around 40 developers has integrated two AI coding assistants, Cursor and Claude Code, into its workflow. The company gave developers the option to choose either tool, and a recent informal survey showed that about 8 developers rely heavily on these agents for most of their coding, while another 11 alternate between using them and traditional methods, with each agent being useful about half the time. Claude Code vs. Cursor Claude Code and Cursor serve different purposes and have distinct strengths. Claude Code, especially with newer models, is adept at writing entire features and prefers to make a plan before executing code. This approach ensures a more structured and comprehensive solution, but it can sometimes be overly ambitious, leading to incomplete or incorrect fixes. Developers must remain vigilant to avoid being misled by seemingly perfect draft pull requests (PRs) that hide underlying issues. Cursor, on the other hand, focuses on making smaller, more contained changes. It integrates directly into developers' editors, making it more accessible and reducing the likelihood of forgetting to use it. While it excels at straightforward tasks, it might struggle with more complex problems, where critical thinking and a deep understanding of the codebase are essential. Impact on Developer Productivity According to the company's estimates, the use of these AI agents has increased developer productivity by about 20%. However, this gain is not uniform across all tasks. Agents shine in handling repetitive and mundane chores, such as converting legacy React components to a new design system. One developer, for instance, saved an hour on a particularly tedious refactor that would have taken several hours to complete manually. This not only boosts efficiency but also helps maintain developer morale by reducing the tedium associated with low-value tasks. For more intricate tasks, such as implementing locking mechanisms to prevent concurrent operations, the agents require more guidance and nudging. Initially, Claude Code proposed a custom Redis solution, only to later suggest built-in Rails methods after further input. This highlights the importance of developers retaining control and applying critical judgment when using these tools. Challenges and Considerations Despite the productivity benefits, integrating AI agents into the development process comes with challenges. One major limitation is the tendency for developers to forget to use the tools, especially when Claude Code is accessed via the terminal compared to Cursor, which is always present in the editor. Another challenge is the potential for reduced code elegance. Agents often write verbose and heavily commented code, which can be functional but lacks the elegance and architectural refinement that humans bring. Overwriting these comments can be time-consuming, but the speed at which agents produce initial drafts can still be valuable. Additionally, the use of agents can homogenize coding styles, making it harder to identify individual contributions. This loss of personal touch might be a temporary concern, akin to the transition from assembly to higher-level languages, but it is noteworthy for those who value unique coding approaches. Company Perspective The company has not mandated the use of AI agents, opting instead to let developers adopt them voluntarily based on perceived benefits. The Chief Financial Officer (CFO) is willing to allocate up to $100 per developer per month to support this technology, considering the potential 20% productivity increase a worthwhile investment. This reflects the company's forward-thinking approach and belief in the transformative power of AI in software development. Industry Insights and Future Trends Industry insiders are excited about the potential of AI in coding, but also cautious. Being proficient at using these agents is emerging as a critical skill, with the ability to craft effective prompts potentially becoming as valuable as traditional coding expertise. Even with advanced agents, the hardest aspect of programming remains defining the software's requirements and articulating them clearly. Syntax and implementation are becoming increasingly automated, allowing developers to focus more on high-level design and problem-solving. Cursor, with its fixed-price model, appears to be designed to encourage frequent use without breaking the bank. Conversely, Claude Code's pay-per-token model suggests a more flexible cost structure, though heavy users might still end up spending a significant amount. Both models highlight the growing trend of AI-assisted development and the need for companies to adapt their workflows and budgets accordingly. In conclusion, while AI coding assistants are powerful tools that can enhance productivity and handle mundane tasks, they require careful integration and continuous human oversight to ensure the highest quality and most effective outcomes. As the technology improves, developers who master the art of prompting and critically evaluating AI-generated code will be well-positioned for success in the evolving landscape of software development.