Escalate difficult decisions to an advisor model

Goal

Let a regular agent handle tasks and consult a stronger model only on explicit triggers, retaining the rationale.

Steps

1. Define measurable escalation triggers such as low confidence or conflicting evidence, and limit advisor context. [source12]
2. Have the primary agent send the question, attempted steps, and evidence to the advisor; log returned advice as an auditable event. [source12]
3. Compare answer quality and call cost before and after escalation; validate advice in the primary flow and do not grant tool authority. [source12]

How to verify

Test predefined escalation conditions and verify the stronger model receives the task, verified evidence, and escalation rationale.

Troubleshooting

If escalation is missed or overused, adjust explicit triggers and replay the same test set; record model disagreement and label advice as advice.