Benzi – A harness reading LESS source code to write better quality code (github.com)

Benzi, a new AI coding agent, has emerged as a game-changer in software development by eschewing traditional methods of parsing code. Unlike other AI tools that simply load code repositories to respond to queries, Benzi utilizes a real compiler based on tree-sitter to parse project files into a structured, queryable index before addressing user questions. This innovative approach allows Benzi to efficiently track symbols, references, and call flows across the codebase, enhancing the accuracy and quality of its responses. With support for multiple programming languages and an impressive accuracy rate of 78.2% on real GitHub issues, Benzi delivers answers in constant time, significantly reducing source code reading overhead. The significance of Benzi for the AI/ML community lies in its new methodology that improves the efficiency of bug tracking and code editing. By leveraging a lightweight index and a dual-engine structure, it can connect function calls and manage dependencies more intelligently. This streamlines the debugging process, minimizing costs to around a cent per resolved bug, compared to higher expenses with traditional models. Benzi’s unique features, such as persistent memory and real-time tracing during execution, mark a substantial leap in the capability of coding assistants, positioning it as an essential tool for developers seeking to enhance their coding efficiency and accuracy.