Assign review requirements by an action’s potential consequences, not by the agent’s reasoning. Classify tools in the policy engine or tool registry: read-only actions can run freely, reversible operations need lightweight confirmation, and high-consequence actions require approval on every call. section 4.8.8
Match the control to the action:
- Use a confirmation dialog for non-agentic assistants. section 4.8.8
- Require approval for each high-consequence tool call; calibrate thresholds so reviewers do not rubber-stamp requests. section 4.8.8
- Prefer framework-native flows, such as LangGraph interrupts or AWS Bedrock human-in-the-loop options, over custom approval workflows. section 4.8.8
