Generating edge-case scenarios is essential for training, testing and assuring the safety of autonomous systems. It reveals hazardous failures, boosts reliability and ensures the system can operate safely at the limits of its operational design domain. Recent advances in large language models (LLMs) have expanded their generation and analysis capabilities, and several works have explored LLM-driven scenario creation. Most of these approaches still rely on human-in-the-loop or rule-based methods, which restricts coverage of the long-tail distribution. We present a laptop-based local LLM agentic AI framework that autonomously generates adaptive scenarios. The solution is inexpensive, fully controllable and preserves data privacy. It follows a think-act-observe-update-continue loop and exchanges information via lightweight JSON. System requirements, description and environment details are stored as knowledge bases and injected into contextual prompts to ensure realistic scenarios. A second, independent LLM critiques each scenario against domain-specific criteria and suggests refinements. Refined scenarios are executed in simulation, and evaluation metrics are collected. Metric scores are analysed to either exploit similar cases for edge-case generation or explore novel scenarios. Post-hoc clustering of metric scores together with thematic analysis of the rich-text scenario descriptions pinpoints functional gaps and their triggering conditions, providing traceable evidence for safety assurance. Experiments demonstrate that the local agentic AI framework can automate edge-case generation and produce verifiable safety-case artifacts.
Review