An AI agent has been developing production planning software for a factory for five months, replacing a complex Excel-based system. The project utilizes four key control mechanisms: a set of process rules, behavioral tests acting as contracts, a ratchet system to prevent regression, and a living technical specification that expands with operational experience.
The process rules require the agent to run the full CI suite rather than individual files and to update executable specifications alongside code. The system includes a registry of 13 invariants for the planner, with each rule verified by at least two tests. The ratchet mechanism prevents tests from being silently removed or altered, requiring a specific PR description to bypass.
The technical documentation grew from 1,300 to 10,500 lines over five months. To manage this volume, the documentation is decomposed into modular files, allowing the agent to read only the relevant sections for each task. The system relies on a core backend with 26 methods, which simplifies testing by providing a consistent API for data operations.
The project team analyzed 1,246 tasks over three months to compare actual production against the planned schedule. This analysis revealed that machine maintenance schedules, previously ignored in Excel, must be manually tracked to ensure accurate planning. The current setup does not cover security or access control, as behavioral tests focus solely on production logic.
