Three pre-recorded agent runs on real tasks. Watch the loop unfold — thought, tool call, observation, repeat. Each step is a model call. Every observation feeds back as context for the next decision. Three scenarios: a clean run that succeeds in 6 steps; a loop that wastes turns searching the same way; a run that hits the step limit before finishing. The visceral "this is what an agent looks like" — most agent demos hide this; the patterns of success and failure live entirely in the trace.