Guarded runtime

Overview

The runtime should call Guard immediately before a tool executes, using the verified identity and exact proposed action. Earlier planning approval does not replace execution-time enforcement.

• Normalize the action and target.
• Verify identity and permission.
• Evaluate organization policy and risk.
• Allow, deny, or require approval.
• Record the decision and resulting action evidence.

Production checklist

• Keep credentials in a server-side secret store.
• Set explicit cost, latency, privacy, provider, and regional constraints.
• Capture route and evaluation evidence for incident review.
• Test timeouts, cancellation, fallback, and denied-policy paths before rollout.