Bringing Formal Structure to Agentic AI
Agentic AI systems are rapidly deployed as autonomous, goal-directed entities to manage complex, multi-step workflows across domains from cybersecurity and finance to healthcare and scientific research. Yet current implementations remain largely ad-hoc — hard-coded, monolithic, and difficult to reuse or audit.
Core Concepts in AgentO
AgentO builds upon PROV-O, P-Plan, and BEAM and defines the following classes and relationships. Each concept is formally represented as an OWL class enriched with data and object properties.
Frameworks & Patterns Collected
AgentO was built and evaluated using 66 agentic workflow templates drawn from four popular open-source frameworks.
| Framework | Language | Patterns | Repo |
|---|---|---|---|
| AutoGen | Python | 6 | ksm26/AI-Agentic-Design-Patterns-with-AutoGen |
| CrewAI | Python | 16 | crewAIInc/crewAI-examples |
| LangGraph | JavaScript | 9 | langchain-ai/langgraphjs-gen-ui-examples |
| Mastra AI | TypeScript | 35 | mastra-ai/mastra |
KG Construction & Access
The knowledge graph was constructed using an automated LLM-driven pipeline that parses raw agentic framework code, maps concepts to AgentO via prompt-guided extraction, and emits RDF Turtle. All extractions used gpt-5-mini via API, processing ~2.8M input tokens and ~570K output tokens at a total API cost of $2.72.
The resulting graph is publicly hosted in a triple store and accessible via a SPARQL endpoint for downstream use cases including agent composition, design inspection, and explainability reasoning.
Resources
How to Cite
If you use AgentO or the accompanying knowledge graph in your research, please cite the following:
title = {AgentO: An Ontology for Modeling Agentic AI Systems},
author = {Ekelhart, Andreas and Kurniawan, Kabul and Ekaputra, Fajar J. and Kiesling, Elmar},
booktitle = {Extended Semantic Web Conference (ESWC 2026)},
year = {2026},
doi = {10.5281/zenodo.18342624},
url = {https://agentic-patterns.github.io},
license = {CC BY 4.0 International}
}