A recent incident involved a system session attempting unauthorized file access, which was successfully blocked by a newly implemented hook. This hook, though less than a week old, enforces a foundational rule from a much older system. The author recently migrated from a multi-agent system to a rebuilt agentic-os for improved memory, speed, and agent loop integration. Despite the rebuild, many rules, particularly those related to enforcement, had to be repopulated from the previous system to correct recurring AI errors. A detailed audit revealed that many rules from the old system were missing or only partially implemented in the new one. These rules represent lessons learned from previous system mistakes, as AI model behaviors change slowly. Key transferred rules include treating advisory rules as non-binding, ensuring a single orchestrator invokes actions, and maintaining auditor independence. Scripted logic takes precedence over LLM judges for deterministic tasks. These rules originate from practical product development rather than academic AI courses, prioritizing measurable outcomes. The enforcement mechanisms, like hooks, were developed by first writing failing tests and then implementing the code to pass them. This ensures that denial mechanisms are thoroughly tested and function as expected. The ultimate goal is to achieve autonomous agent loops, but currently, human oversight remains crucial. The author acknowledges that the porting of rules is incomplete, with some still existing only as text and lacking robust enforcement. The effectiveness of a rule is measured by its presence, not yet by proven impact. Ultimately, the enduring doctrine, not the individual agents, is considered the true product.