Where Should Intelligence Live? Agent-Centric vs. Environment-Centric Design in Agentic Systems

When building agentic systems, practitioners face a critical choice: should intelligence live in versatile agents that adapt to any environment, or in specialized systems designed for specific tasks? This question manifests vividly in robotics-would you rather have a humanoid robot drive your car, or a purpose-built self-driving system like Waymo? The answer seems obvious, yet the same question applied to software divides the AI engineering community. This paper provides a practical framework for making that architectural decision. We show that specialized systems dominate in structured environments where reliability is paramount, while general agents suit unstructured settings where infrastructure modification is impractical. We analyze failure modes unique to each approach-including context rot, hallucination cascades, and needle-in-haystack retrieval failures-and argue for hybrid architectures analogous to cognitive architectures with an LLM brain orchestrating specialized tools. The framework helps practitioners design robust agentic systems by matching architecture to task diversity, environment controllability, and safety requirements, while acknowledging that rapid AI progress may shift these trade-offs significantly within the decade.