Bugbot is a code review agent designed to catch logic bugs, performance issues, and security vulnerabilities before code reaches production. Initially, the quality of code reviews was poor due to limitations in the models used. Over time, the team improved Bugbot's effectiveness through systematic experimentation, increasing its resolution rate from 52% to over 70%. They managed to double the average number of resolved bugs flagged per pull request, from 0.2 to 0.5, by refining the review process and using methods like parallel bug-finding passes and majority voting to filter results. To enhance its functionality, Bugbot underwent a significant overhaul by switching to an agentic architecture. This allowed the system to reason through code changes and adapt its approach dynamically, rather than following a fixed sequence. The team also developed a resolution rate metric to measure Bugbot's performance effectively, which provided clearer insights into its impact on code quality. With over two million pull requests reviewed monthly for various clients, the tool is now significantly more effective than it was at launch, and the team is actively working on new features like Bugbot Autofix, which automates bug fixes during reviews. The evolution of Bugbot shows a clear trajectory of improvement driven by experimentation and feedback from internal engineers. Key developments included enhancing repository access, adding custom rules for specific codebase checks, and refining the model's behavior based on new insights. The ongoing focus is on leveraging emerging models and advancing the system's capabilities to support scalable AI development workflows.