THE SAFETY ARCHITECTURE

RUNTIME APPROVAL SYSTEMS WITH PERSISTENT PERMISSIONS Some actions are too risky to allow automatically. Runtime approval means that when the agent wants to take such an action, the harness pauses and asks a human or a policy engine for a decision at that moment. Typical prompts say "The agent wants to run this command. Allow once, allow always, or deny?" PERSISTENT PERMISSIONS If the system asked about every tiny action, people would become annoyed and start clicking allow without reading. This is called approval fatigue, and it silently destroys safety. Persistent permissions solve part of the problem. When a user approves "this agent may always read files in this folder," the harness stores that decision and does not ask again for the same kind of action in that scope. But it still asks for a new and different kind of action, such as deleting files or contacting the internet. Good approval systems are specific about scope: which tool, which resource, how long, and for whom. They show the human clear information, such as the exact command, not a vague summary. They allow permissions to be revoked. And they record every decision in a log for later review.