Knowledge design in complex domains

Abstract

Traditional Knowledge Engineering (KE) often fails to address modern complex domains characterized by "wicked problems," where non-stationary ground truths and high dimensionality render static rule sets obsolete. This paper proposes a paradigm shift from traditional Knowledge Engineering to a co-agentic Knowledge Design model. While classical KE relies on static, human-led abstractions that often fail in dynamic environments, we argue for a model of co-agency where human domain expertise and Large Language Model (LLM) generative capabilities are distributed across a strategic Knowledge Level and an operational Application Level. Through the proposed AI-in-the-Loop (AIL) architecture, agency is shared: the human expert provides strategic governance and critical plausibility review, while the AI autonomously operationalizes these directions into a dynamic Chain-of-Agents (CoA). This co-agentic synergy enables the rapid creation of Tailored Expert Systems (TES), minimalist case-specific reasoning paths that co-evolve with the problem space in real-time. The framework is demonstrated in the facility management domain through fire safety planning, utilizing the Cocoanut Grove disaster as a complex testbed. Experimental results demonstrate that the system exhibits high structural stability and sensitivity to both regulatory constraints and real-world operational changes. Expert review confirms the plausibility and internal consistency of the generated recommendations. While current limitations include the need for refined human-AI interfaces and formalized evaluation metrics, the AIL framework offers a robust blueprint for the next generation of adaptive and auditable expert systems.