Semantic Modeling for Auditable, Reusable, and Traceable Agentic AI Systems

Agentic AI systems — comprising autonomous agents capable of planning, tool use, reflection, and collaboration — are increasingly adopted to automate complex, multi-step processes. In cybersecurity, such systems promise to streamline threat detection, analysis, and response by orchestrating heterogeneous tools and knowledge sources. However, current agentic AI implementations remain largely ad-hoc and opaque, relying on imperative code and prompt-based configurations that are difficult to reuse, validate, explain, or audit. This lack of formal structure raises significant issues in high-stakes environments, where transparency and accountability are essential.
The proposed project Semantic Modeling for Auditable, Reusable, and Traceable Agentic AI Systems (SMART-AAI) aims to introduce a declarative, knowledge-graph-based foundation for engineering agentic workflows. The core objective is to transform agentic AI from black-box pipelines into structured, transparent, and reusable systems by explicitly modeling agents, tasks, goals, tools, resources, workflow patterns, and execution traces using Semantic Web technologies. Building on prior work in agentic AI ontologies and cybersecurity knowledge graphs, the project will develop an integrated semantic framework that supports design-time validation, runtime provenance, and post-hoc explainability of agentic behavior.
The project will construct an expressive ontology and an accompanying Agentic AI Knowledge Graph, enriched with formal constraints using OWL and SHACL to validate workflow correctness. Runtime execution events — including agent decisions, tool usage, and task completion — will be captured as provenance-aware knowledge graph updates, enabling detailed execution reconstruction and auditing. A declarative-to-executable translation layer will bridge semantic specifications with existing agentic AI frameworks, allowing practitioners to define workflows declaratively while preserving practical deployability.
Cybersecurity will provide an excellent validation domain for the developed framework; specifically, we will focus on incidence response scenarios and use the developed semantic methods to construct agentic pipelines that integrate threat intelligence, attack knowledge graphs, and knowledge-graph-enhanced retrieval. The project will evaluate expressivity, reusability, and explainability through quantitative experiments and failure-injection studies. By delivering open-source tools, reusable agentic workflows, and rigorous empirical results, this project establishes foundational methods for trustworthy agentic AI and advances the state of the art in declarative agentic systems engineering.