I recently heard the CEOs of DeepMind and Anthropic discussing this thought experiment: Imagine a new country appearing overnight with a population of 10 million people, all smarter than Albert Einstein. What would happen? In another podcast, Microsoft’s CEO described a future where a company might have 2 million or even 20 million AI agents. What many people don’t realize is that an “AI agent” is really just a coordinated series of steps. For example, managing household trash could involve: 1. One agent tracking the garbage can’s location at all times. 2. Another monitoring the city of Cary’s website for schedule changes. 3. A third setting alerts if anything updates. 4. A fourth sending a reminder text to a human or robot on trash day. These can be stacked endlessly into complex systems. When people say “AI,” they often mean this full harness—the agents, sub-agents, and rules that tie them together. Once we have truly capable agents, they will design and optimize these systems themselves: creating the plans, fallback procedures, and actions (using local/personal LLMs or SLMs). The next phase—recursive self-improvement—is already close. After that, progress becomes unstoppable and life gets very strange.