ATLAS: Agentic Traversal for Large-scale Analysis of Systems

A Python-based experimental framework designed to benchmark Agentic Navigation against standard Retrieval-Augmented Generation (RAG) for code repository understanding. Instead of relying solely on semantic vector search, ATLAS parses Python repositories into a dependency graph (using AST and NetworkX) and allows an LLM agent to traverse the graph node-by-node to locate the root cause of issues. The repository includes: The ATLAS Agent: A traversal agent using OpenAI's API. Graph Builder: Tools to parse imports and build directed graphs from codebases (Flask, Django, SymPy). Statistical Analysis: Scripts to perform McNemar's Test (accuracy) and Paired T-Tests (efficiency) comparing the Agent against a RAG baseline. Benchmark Data: Real-world issue reproduction cases.