Scaling AI agents beyond demos is fundamentally a systems design problem, not a model capability problem. As a single agent's scope grows, costs rise non-linearly, failures propagate across the entire execution chain, and context becomes noisy and hard to manage. The solution is decomposing monolithic agents into multi-agent systems with bounded, distributed responsibilities. This introduces a new trade-off between horizontal scaling (adding specialized agents, increasing coordination overhead) and vertical scaling (embedding capabilities within existing agents, increasing per-agent complexity). A practical rule of thumb: split capabilities when they are reusable and independent, embed them when tightly coupled and context-dependent. Successful agentic systems are those that deliberately choose where complexity accumulates.