[storystarling]

True, but that's still 1,500 inference cycles. Even without external API fees, the latency and compute burden seems huge. I don't see how the economics work there without significant subsidies.

[darrinm]

FWIW many tool calls can be and often are made in one inference cycle.