Prevent unauthorized AI agent actions

Kastra evaluates integrated agent actions against policy before execution. Unauthorized actions are blocked, redacted, or placed on HOLD for human approval.