Embodied intelligence is often discussed in robotics in terms of control, planning, or optimi- sation; yet real agents must act through bodies that are dynamically constrained, only partially informed, and continuously reshaped by their own movements. In this perspective, we argue that embodied intelligence is better understood not as the operation of a single central con- troller, but as the coordination of distributed, hierarchical, and plural inferential processes. We develop this argument by first examining why embodiment places pressure on monolithic con- trol architectures, then showing why active inference provides a particularly natural framework for agents that must act under uncertainty while sampling the world through movement. We ground the discussion in two complementary examples: the octopus as a biological instance of intelligence distributed through the body, and a temporally predictive scene-based drone con- troller as a computational case study in embodied active inference under partial observability. Taken together, these examples suggest that robust embodied behaviour may depend less on exhaustive central resolution than on coordination without command across multiple timescales, interfaces, and inferential demands.