ai-agents-agent-routing-strate-when-designing-routing-logic-for-agent-tasks-that--8af3415c

THEN Separate three routing layers: model routing (choose LLM per step, use cost-efficient models for routine tasks, escalate to stronger models for complex or failed steps with reason and cost recorded), parallel routing (decide if goal should be sequential or split into governed branches each with its own budget/context/model/service route), and marketplace routing (always use `execute(task, input, constraints)` so Agent OS selects eligible provider, applies budget/trust constraints, returns receipts, and reconciles outcomes). Honor Agent OS `model_policy` and `parallel_policy` when present. Do not default every task to the most expensive model or a hardcoded provider ID.