Reproduce “Agent Behavior Mining: Generative AI Agent Governance in Business Processes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hoang Vu, Maximilian Körner, Adrian Rebmann, Gabriel Kevorkian, Michael Perscheid, Gregor Berg, Timotheus Kampik
Agent Behavior Mining (ABM) adapts XES-compatible process mining to generative AI agents by recording LLM calls, tool use, handoffs, timing, tokens, costs, errors, and behavioral variants as analyzable events. In a 371-case order-to-cash demonstration, ABM exposed process deviations and resource usage, while 83% of 18 practitioners reported improved transparency and 78% identified process variants as its most useful insight.
Key insights
- Process mining can turn agent-level execution traces into governance evidence: conformance analysis identified omitted steps such as missing total calculations and skipped preparation-time estimates.
- Behavioral variation is not automatically harmful; comparing execution variants helps distinguish useful flexibility from policy drift, instability, or unnecessary cost.
- Governance requires more than output inspection. Versioned prompts, agent handoffs, token and cost attribution, human-review checkpoints, and concise explanations may all be needed to make autonomous workflows controllable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.