Why a robot can't think for a few seconds
A text-based chat agent can often take a few seconds to reason before responding, and a user will barely notice. A physical agent usually doesn't have that luxury. If someone waves at a robot, speaks to it, or steps into its path, a multi-second pause before it reacts doesn't just feel unnatural, it can be unsafe, since the world keeps changing during that pause. Responding to voice and visual input with low latency is what makes a physical agent feel present and trustworthy instead of laggy and, worse, late to notice something it should have reacted to immediately.
This pushes physical agent architectures toward techniques that trade some reasoning depth for speed where it matters most. Simple, fast reflexes, like stopping when something enters an arm's path, are often handled by lower-level, low-latency logic rather than waiting on a full pass through the LLM-based reasoning engine, while the slower, more deliberate agent loop handles higher-level planning that can tolerate a bit more delay.