Uwe Zdun

Reinforcement Learning (RL) systems are gaining traction in application areas such as robotic systems, self-driving vehicles, and automated industrial operations.In these safety-critical domains, lifecycle management practices – including model versioning, registry integration, automated evaluation, retraining triggers, and staged deployments – are just as important as algorithmic performance in ensuring reliability and reproducibility. However, these practices are often scattered across source code, configuration files, CI/CD workflows, and deployment manifests, making it difficult to automate detection and correction. This paper introduces an Agentic AI architecture that utilizes multiple specialized LLM-based agents to evaluate RL pipelines against a curated catalog of 15 architectural best practices. The agents are specialized (e.g., file selection, detection, fix suggestion, code generation), and an orchestrated pipeline with automated validation gates combines practice detection and LLM reasoning to identify both explicit and implicit practice gaps and suggest, or where appropriate, even apply fixes. We validate the approach in a large-scale industrial cyber-physical production case study and in five open-source RL repositories with diverse architectures, demonstrating the scalable detection of missing practices and the generation of actionable fixes. The system achieved an overall detection F₁-score of 0.80 and a mean generation-quality score of 2.40, indicating that it reliably identifies missing practices and produces functionally valid code artifacts.