Why our agents run slower than they could
Latency we could remove and choose not to, because a cheaper model that waits beats a fast one that guesses.
What 4,812 runs taught us about cost per run
Four thousand runs of billing data, and the three things that actually moved cost per run.