Toward Safe Intelligence: Non-Agentic Architectures as a Structural Alternative to Objective-Based AI Alignment

This preprint introduces non-agentic architectures as a structural alternative to objective-based AI alignment. It argues that systems without internal objectives, temporal projection, or narrative continuity exhibit intrinsically safer computational dynamics than agentic optimizers. The paper analyzes the structural fragility of agentic systems, outlines the principles of non-agenticity, and proposes a two-phase research program: (1) an experimental simulation using current large language models with externalized memory and non-decisional, resonance-based selection mechanisms; and (2) the development of native non-agentic architectures grounded in present-state computation. The approach aims to provide a viable pathway toward intrinsically safe intelligent systems by avoiding the emergence of instrumental drives and self-preservational behaviors.