the overnight cost thing is real. "$200 in 3 days" is actually pretty tame compared to what happens when you have agents spawning sub-tasks without a budget cap.

the part that doesn't get talked about enough: most people are hitting a single provider API and treating it as fixed cost. but inference pricing varies a lot across providers for the same model. we've seen 3-5x spreads for equivalent quality on commodity models.

so half the cost problem is architectural (don't let agents spin unboundedly) and the other half is just... shopping around. not glamorous but real.