Multi-Agent Auditing for Smart Contracts*

Abstract

Smart contracts underpin contemporary decentralized systems, yet their immutability and perpetual execution amplify the consequences of latent defects. Despite progress in manual audits, static analysis, fuzzing, and formal verification, auditors face a widening gap between the scalability and desired assurance due to limited automation capability. Recent large language models (LLMs) with tool-use capabilities promise greater automation, but monolithic single-agent auditors struggle with coverage, robustness, and reproducibility. Motivated by addressing these issues, we propose Multi-Agent Auditing (MAA), a framework that coordinates a team of tool-grounded agents through a constrained protocol that privileges verifiable artifacts. Besides, we mechanize an LLM-assisted orchestration mechanism and a shared knowledge base to coordinate a set of agents specialized in sophisticated testing approaches to produce budget-aware and evidence-centric audit results. Furthermore, we present the experiment results showing that MAA outperforms singleLLM auditors and provide empirical insights into LLM-backend selection.