Most AI agent observability tools (Langfuse, LangSmith, etc.) show you what an agent did after the fact. I wanted something that actually stops a risky action before it runs. Built WorkerZero — you register an agent, define policies (like "require approval if amount > $500" or "block account deletions"), and every proposed action gets evaluated in real time before it executes. Allow / block / require human approval. It's a real working MVP — not a mockup. Live demo.