From Code to Knowledge: Enhancing Coding Agents via Reverse-Engineered Software Knowledge

Abstract

Despite advances in code retrieval and static analysis, LLM-based coding agents rarely construct or consult structured representations of a codebase's architectural design, domain model, or behavioral patterns, the very knowledge that human developers depend on when making nontrivial changes. This absence of explicit knowledge models may cause agents to produce code that, while syntactically correct, can violate architectural constraints, ignore established patterns, and fail to align with the codebase's conceptual vocabulary. We argue that reverse-engineering this structured knowledge from existing code and injecting it into coding agents addresses this limitation. We propose a three-layer knowledge taxonomy covering architectural, domain, and behavioral knowledge; a multi-level extraction framework using static analysis, schema mining, and pattern detection; and integration strategies based on prompt augmentation and retrieval-augmented generation. Our approach is lightweight, model-agnostic, and can layer onto existing coding agents without retraining.