Claude Code built PM4Py-UCM process-mining tool with 691 tests, 151 commits
18 agent sessions, 10,328 tool actions, and a 2.3:1 fixes-to-features ratio
Enterprise-modeling tools are notoriously complex to extend, but AI coding agents are now stepping in to build entire modeling utilities from scratch. In a new arXiv paper, Daniel Amyot documents the AI-assisted development of PM4Py-UCM, an open-source tool that mines Use Case Map (UCM) models from event logs—adding process-mining capabilities like performance heatmaps, dashboards, and executable scenario variants.
Claude Code powered the build through 18 agent sessions spanning 374 human turns and 10,328 tool actions over 65 hours. The development produced 151 commits and 20 releases, expanding the test suite from 108 to 691 test functions. Notably, fixes outnumbered features 2.3:1 up to version 0.7.4, with roughly 18% of human turns dedicated to correcting agent errors. Cross-cutting consistency work—docs, tests, notebooks—lagged feature waves, highlighting a key challenge for AI-driven tooling. The research offers lessons on making model transformations mechanically checkable and using oracle-based validation to close the "the agent said it works" gap.
- PM4Py-UCM mines Use Case Map models from event logs with heatmaps and dashboards
- Built with Claude Code: 18 sessions, 151 commits, 20 releases, 10,328 tool actions
- Test suite grew from 108 to 691 functions; fixes outpaced features 2.3:1
Why It Matters
Shows AI agents can ship real modeling tools, but rigorous validation still costs engineering time.