Are you guys also hitting a cost wall with agents? Any harnesses that actually support Batch API?

I’ve been tracking my agentic workflow costs, and I'm realizing a massive chunk of the budget is being leaked because my bg agents are treating everything as "realtime" inference. Has anyone found an agent harness or orchestration pattern that handles this better? Are there any frameworks that treat "Batch API" as a first-class citizen in the agent loop? Or are most of you building custom queuing/buffer layers to group requests before hitting the models? I’m currently debating whether to build a custom batch-native orchestration layer, but I’d rather not reinvent the wheel if there’s a pattern or library I’m overlooking. Would love to hear how you’re keeping agent costs down in production, especially for tasks that don’t require an immediate human-in-the-loop response.