Inference has spent years getting better at generating tokens.
But agents don't get paid to generate tokens. They get paid to finish jobs.
Optimizing individual requests won't solve the whole workflow.
@pasha_aqil talks about why inference needs a rethink:
