Greptile Co-founder Daksh Gupta: How to Build an Effective AI Code Review Agent
Greptile co-founder Daksh Gupta shares engineering practices for building AI code review agents. He analyzes the limitations of simple sandbox-plus-webhook setups (missing bugs, high costs, noise) and details six fixes, including targeted prompts, historical incident PR datasets, model inversion (using GPT to review Claude-written code and vice versa, catching ~10% more bugs), codebase context with a living wiki, autonomous testing, and routing cheaper models to most subtasks.
For engineering teams looking to build internal AI code review tools, adopting a simple 'sandbox + agent + webhook' architecture often leads to three major pain points: missing critical bugs, high per-review costs ($3-$25), and flooding developers with nitpicky comments. In his latest video, Greptile co-founder Daksh Gupta systematically deconstructs these issues and shares six improvements validated through Greptile's actual operations.
Gupta highlights that improving review quality hinges on refined context and strategy. First, use more precise prompts to reduce noise and leverage historical datasets containing past incident PRs to train models on identifying real risks. Second, he emphasizes 'Model Inversion,' a technique where GPT reviews code written by Claude, or Claude reviews code generated by GPT. According to his description, this cross-review mechanism catches approximately 10% more bugs. Additionally, incorporating codebase context and a continuously updated living wiki helps the agent understand project-specific norms and dependencies.
Regarding cost control and execution efficiency, Gupta recommends running autonomous tests in sandboxes using computer use capabilities to ensure the practical viability of code changes. He also suggests implementing smart routing strategies, assigning most routine subtasks to cheaper models while reserving high-performance models for complex segments. This combination aims to balance review depth with operational costs, providing a reference framework for engineering teams.