A real case from the WixQA runs. Each column is one agent turn, and every tool call takes one. Averages are over all questions in each case.
turn 1turn 2turn 3turn 4turn 5
Without a cache3 tool calls here3.2 on average
search9 sections, mostly about mobile
readResizing a gallery with the stretch handles
Cache hit1 tool call here1.2 on average
lookupmatch 0.97, 5 sections handed over
Cache miss4 tool calls here3.4 on average, 3.5 without a cache
lookupnothing found+1 call