Object-Centric Process Mining for AI-Agent Execution Intelligence

Abstract This study investigates the application of Object-Centric Process Mining (OCPM) to the analysis of AI-agent execution lifecycles using Celonis. Drawing on 720 execution runs from the WildClawBench benchmark, the research examines 35,771 execution events and reconstructs 17,206 tool invocations to investigate the relationships among agent activities, tool requests, background processes, and observable failure indicators. The study develops a lifecycle-aware object-centric representation that supports the analysis of execution relationships extending beyond conventional single-case event sequences. The resulting analysis identifies 1,487 batched tool requests and 1,142 linked background-process sessions, demonstrating how object relationships and preserved execution metadata can support more detailed investigation of agent behavior. A comparison of event-log representations shows that 615 of 1,212 rule-classified failure indicators were visible in a simplified representation, while enriched representations retained the information necessary to identify all indicators under the same classification rules. The findings highlight the potential of Celonis OCPM as an analytical framework for investigating AI-agent execution structures, reconstructing tool and process lifecycles, and supporting operational observability. The research contributes an empirical foundation for applying process intelligence to increasingly complex AI-agent systems while distinguishing structural observability from independently validated failure prediction.