Policy-Enforced Execution#

A runtime model where every system operation an agent performs is authorized against an explicit policy before execution. The policy is defined by the host/service owner, separate from the script or agent that requests the operation.

Key Properties#

  • Separation of concerns: script says what to do, policy says what’s allowed
  • Host-owner control: the system owner defines boundaries, not the agent author
  • Graceful denial: policy violations return clear errors — agents can observe, reason, and adapt
  • Defense in depth: even if an agent is compromised (prompt injection, hallucination), the policy boundary holds

Implementations#

ToolLayerMechanism
rexSystem callCedar policy evaluated per operation (read/write/open)
claude-codeTool usePermission modes (ask/auto-accept/plan) per tool category
kiroTool approvalPer-tool permission prompts, pre-approval in agent config

Rex operates at the lowest layer — individual system calls. Claude Code and Kiro operate at the tool-invocation layer. These are complementary, not competing.

Relationship to Governance#

This concept makes agentic-ai-governance pillar 2 (least privilege) and pillar 5 (kill switches) concrete and enforceable at runtime. It’s the technical implementation of what governance frameworks describe in policy terms.

See Also#