Abstract
Static analysis remains crucial for ensuring software reliability and security, yet it frequently struggles with false positives, rigid rule sets, and an inability to reason about semantic context. LLM-based agents show promise as a complement to traditional approaches by bringing autonomous reasoning and verification capabilities to the analysis pipeline. This paper introduces a taxonomy that organizes these agents along four key dimensions: architecture, integration patterns, grounding mechanisms, and application domains. Our framework captures how agents are structured, where they intervene in existing workflows, how they validate their outputs, and which tasks they address - from defect detection and code maintenance to security auditing. We examine current evaluation practices and surface major limitations around reliability and scalability. The paper concludes by distilling lessons from recent work, highlighting open research challenges, and identifying promising directions for future investigation.