Time-non-critical agents are super slow because you use flex mode in OpenAI to save costs. Now with gpt-5 they became even slower (and cheaper, so it's a good thing!). So you *need* to parallelize your agents. This is what I've been doing last night. My library #monoflux's flatMap method on a Flux (a Flux is an asyncronous array) can now handle () => Flux callbacks. It is now my number 1 engine to run *tons* of agents in parallel.