The Shadow Sandbox: Where an AI Agent Earns Its Permissions Before It Touches Anything Real

Before an AI agent touches a real system, it should prove itself somewhere safe. I built a sandbox for that: a simulated factory where an agent handles customer orders, and everything it does is measured against what the human order desk did. At first the agent only watches and decides in the background, without acting. When it agrees with the humans often enough, it earns the next level: suggesting, then proposing plans that a person approves, then acting on its own where actions can be undone. Actions that cannot be undone, like confirming an order to a customer, always stay with a person. When people correct the agent, it remembers the correction, turns repeated ones into rules, and drops back a level until it has proven itself again. The sandbox follows a set of company rules that the agent is never shown, so I can check exactly which ones it learned, and halfway through the run the rules change. I compare the results with an agent that never learns, and the talk includes a live demo, the numbers, and the open-source code.