Picture an agent answering a support question. One step is one pass of the loop: the model reads the context, picks a tool, the tool runs, and the result is added to the context. With a 2k-token context, one step takes about 4 seconds and usually works.
Nobody worries about one step. The trouble is that real tasks need many of them, and three costs grow as the run gets longer: tokens (money), time (the user waits), and failures (more chances to go wrong).
✦Key idea
Measure a whole run, not one step. Steps multiply every cost.