At a Glance
A two-agent local system can take plain-language business questions, generate and verify executable analysis code, and produce data-grounded reports without sending raw event logs to the cloud.
ON THIS PAGE
Key Findings
Splitting work between a code-focused agent and an interpretation-focused agent makes process analysis more reliable: the code agent writes Python to compute metrics, a static checker and controlled runner prevent unsafe or flaky behavior, and the report agent turns the outputs into human-friendly findings. Only lightweight schema snapshots (column names and types) are shared with the agents, so raw event data stays local and privacy risks are reduced. In a demo on a real loan application log, the system automatically produced runnable code and complete, visual reports for common business questions. Tool Use Pattern
Data Highlights
1Handled 5 business questions (Q1–Q5) from the demo; the code-generation agent produced executable Python for all 5 tasks (100% success in the demo).
2System uses a two-agent split — an Engineer node for code synthesis and an Analyst node for interpretation — to separate technical and semantic work.
3Demo used the real-world BPI Challenge 2017 loan application dataset (2017), showing end-to-end local execution with no raw data leaving the host.
What This Means
Engineers building AI assistants for business users — because this pattern produces verifiable, runnable analysis instead of verbal summaries. Technical product leaders and data platform teams — because it preserves data governance by keeping raw logs local while still offering a no-code experience. Researchers interested in agent orchestration — because the architecture formalizes a practical, inspectable workflow for autonomous analysis. Role-Based Agent Pattern
Avoid common pitfallsLearn what failures to watch for
Key Figures
Fig 1: Figure 1 : Overview of the proposed architecture.
Fig 2: (a) Code generation process.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
The evaluation is a demonstration on one public dataset and five specific business questions; broader generalization to other domains and larger, messier logs needs further testing. The demo used a high-end language model for generation; deploying fully offline or with smaller models may reduce code quality and require more tuning. Security depends on the static checks and controlled execution environment—these need rigorous hardening and auditing before production use in sensitive settings. Guardrails Pattern
The Details
PMAx rearranges process mining into an agent workflow that trades guessing for deterministic computation. Users upload an event log locally; the system extracts only a lightweight schema snapshot (column names, types, example values) so agents get context without seeing raw data. A specialized Engineer agent synthesizes Python scripts that create process artifacts (models, summary tables, charts). Before any code runs, a static verification layer enforces security and syntax constraints; code executes inside a controlled environment with an automated self-correction loop that captures runtime errors and iteratively refines the scripts. Dynamic Task Routing Pattern
An Analyst agent receives the produced artifacts’ metadata and generates the natural-language report that answers the user's original question, combining textual interpretation with charts and tables. The two-agent divide-and-conquer design keeps technical details out of the Analyst’s context so reasoning remains focused and compact. The framework is open-source, built on common Python tooling, and was demonstrated on the BPI Challenge 2017 loan log: for five practical business queries the system generated executable code and produced data-grounded reports. The approach improves auditability and privacy for conversational process analysis, but needs more benchmarks, tighter production-grade security, and testing with diverse event log shapes before widespread deployment. Multi-Agent Knowledge Management
Need expert guidance?We can help implement this
Credibility Assessment:
Contains a highly recognized author (Wil M. P. van der Aalst), a top researcher in process mining/AI-related fields — strong author reputation despite arXiv venue.