DeveloperWeek 2025

YK Sugi, Sourcegraph, Senior AI Developer Advocate

This presentation details the development of an experimental software engineering agent designed to answer complex questions about codebases. The core innovation lies in a graph-based approach where the agent iteratively gathers contextual information, using self-reflection to assess its confidence in answering the query. By treating accumulated context as a set and employing a greedy best-first search guided by confidence scores, the system aims to efficiently navigate the codebase and provide accurate answers. Initial findings are promising, with future benchmarks on real-world software engineering challenges planned to further validate its capabilities.