How do you manage cost and latency in production agent deployments?