Paper proposes layered architecture for Agentic AI using OpenClaw and Ollama

A new arXiv paper introduces a comprehensive, layered architecture for Agentic AI that explicitly separates inference, orchestration, and execution layers in autonomous agent systems. The authors argue that current reactive LLM interfaces are insufficient for persistent, goal-driven behavior and outline an evolution toward systems with memory, planning, and continuous execution.

The paper analyzes OpenClaw and Ollama as a full-stack Agentic AI system, where Ollama provides the LLM inference layer and OpenClaw handles runtime orchestration, integrating reasoning, tool use, and action execution. The authors report that a prototype implementation demonstrates capabilities such as persistent memory, tool utilization, and adaptive decision-making emerging from system-level integration rather than from standalone models alone.

According to the study, performance in the prototype improved consistently as architectural complexity increased, suggesting that system-level design choices drive measurable gains in agentic behavior. The authors also examine challenges including scalability, security, privacy, governance, and evaluation, emphasizing the need for robust benchmarking and system-level design practices.

The paper concludes by proposing future directions such as scalable multi-agent architectures, distributed autonomous systems, and human-aware Agentic AI frameworks aimed at responsible deployment. All models, code, and datasets from the research are publicly released to support reproducibility and benchmarking.