Ed Huang highlights a significant shift in infrastructure software use, where AI agents are becoming the primary users instead of human developers. He points out that over 90% of new TiDB clusters on TiDB Cloud are created by AI agents, which reflects a broader trend in production environments. This change prompts a reevaluation of how foundational software should be designed. Huang emphasizes that, as AI interacts with software, it relies on established mental models rather than user interfaces or APIs. These mental models are rooted in stable, long-standing concepts like file systems and SQL, which remain relevant even as technology evolves. He argues that designing software for AI agents isn't about creating new interfaces but aligning with existing cognitive structures that AI has already internalized. Huang uses the example of file systems, particularly his experimental agfs, to illustrate that software can be flexible and extensible while maintaining a consistent mental model. He highlights how implementations can add new functionalities without disrupting the user's understanding. This is critical as AI systems evolve rapidly, and stable abstractions help manage that speed without chaos. Huang also addresses the relevance of software ecosystems. While differences in syntax and protocols may seem trivial to AI, they still matter because popular software is often built on well-established mental models. The key takeaway is that if the underlying model is sound, AI can adapt to various systems with ease. This poses a challenge for paradigm-level innovation, suggesting that while AI can navigate existing frameworks, creating fundamentally new ones is becoming increasingly difficult.