I've recently completed an independent research study exploring how Celonis Object-Centric Process Mining (OCPM) can help us understand the execution behavior of AI agents.
The idea was simple: If Celonis can provide deep visibility into complex business processes, could we apply the same principles to understand how AI agents execute tasks, interact with tools, and encounter failures?
To investigate this, I analyzed execution traces from the WildClawBench dataset using Celonis.
Some interesting findings emerged:
-
720 AI-agent execution runs analyzed, covering 35,771 events.
-
17,206 tool invocations reconstructed and examined.
-
1,487 batched tool requests identified, revealing execution structures that are not immediately apparent in conventional event sequences.
-
1,142 background-process sessions linked to their execution context.
-
Failure-indicator visibility increased from 50.7% to 100% when richer execution metadata was preserved, based on our rule-defined classification framework.
What I found particularly interesting was how Celonis OCPM helped represent the relationships between agent runs, tool invocations, and background processes, allowing us to examine execution lifecycles rather than isolated events.
For me, the most promising takeaway is that Celonis could play a valuable role in AI-agent observability, execution transparency, failure investigation, and operational governance as organizations adopt more autonomous AI systems.
I've documented the methodology, implementation, findings, and limitations in a 72-page research paper.
Research paper: Object-Centric Process Mining for AI-Agent Execution Intelligence
Published on Zenodo: https://doi.org/10.5281/zenodo.24030726
I'm curious to hear what others in the Celonis Community think.
Do you see object-centric process mining becoming an important part of monitoring and managing enterprise AI agents?
I'd be glad to exchange ideas with anyone exploring similar applications.
Pratik Sarkar
Independent Researcher