Runtime guardrails for AI agents

• Intercept every tool call before execution
• Allow, block, or route to human approval
• Full audit log with session context