Show HN: AxiomCode Graph, Reducing coding agent rework with formal methods (github.com)

AxiomCode Graph has been introduced as a powerful tool for AI agents, significantly enhancing the understanding of codebases through a formal methods approach. By building a knowledge graph that accurately maps function calls based on logical rules—rather than relying on traditional methods like grepping—it allows AI agents to explore and interact with code more effectively. This meticulous mapping reduces the chances of missing vital connections in complex coding environments, ultimately leading to more accurate task execution and a reduction of up to 50% in tool calls when exploring new codebases. The system boasts impressive accuracy metrics, with an F1 score of 94.9% on bug-revealing tests against held-out Defects4J bugs, outperforming other graph builders like tree-sitter. This is achieved by employing a typed graph that resolves call targets from receiver types, as done by compilers, rather than merely guessing through syntax alone. The engine supports multiple languages, including Java, TypeScript, and Python, with seamless integration into existing agent frameworks. This advancement promises to bolster the efficiency and reliability of AI agents tasked with code modification and debugging, thereby benefiting the broader AI and machine learning community.