First, my own work, using agents to code: right now I am limited by the speed of flagship models, even when running multiple agents in parallel.
Second, the work models do for my clients in the systems I build: mostly quite time-sensitive. One application is an agent-assisted design loop which must be interactive. Another is an automated drafting system that must produce a timely draft.
I always use the batch API when I can, often with a synchronous fallback if the delay is too long. But the number of truly “background” tasks that can wait 24hrs is zero so far.
Hard disagree.
First, my own work, using agents to code: right now I am limited by the speed of flagship models, even when running multiple agents in parallel.
Second, the work models do for my clients in the systems I build: mostly quite time-sensitive. One application is an agent-assisted design loop which must be interactive. Another is an automated drafting system that must produce a timely draft.
I always use the batch API when I can, often with a synchronous fallback if the delay is too long. But the number of truly “background” tasks that can wait 24hrs is zero so far.