Autonomous agents don’t just fail. They persist. They retry, replan, and chain tools until something “works.” That persistence is exactly what makes agents valuable, and exactly what makes them hazardous in production without strict execution controls.