The same question goes down three paths at the same moment, with the same account, the same Work IQ tools and the same model writing. The steps, tool calls and timings are real, from the app’s history. Names and answer text are changed. Bars are drawn to scale on one clock. Played at 4x speed.
Names are changed. On the bottom row Work IQ ask alone was first, by 10s, with the full meeting summary. Jev looked in search first, saw the summary was not in those results, and passed the question to ask, so it ran both steps. The model picking tools replied sooner with a shorter answer (hollow dot). Three runs where my own test rig hit a one-off hiccup are left out.