Core Insights - The main trend observed is the shift in primary users of infrastructure software from human developers to AI agents, with over 90% of new TiDB clusters being created directly by AI agents [1] - This shift challenges traditional assumptions about how databases should be used and necessitates a reevaluation of the essential characteristics that software should possess when designed for AI agents [1] Group 1: Characteristics of Software for AI Agents - The core user of software is transitioning from humans to AI, which means that the underlying mental models exposed to users are no longer UI and API, but rather the mental models behind them [2] - AI agents, having been trained on vast amounts of code and engineering practices, recognize repetitive patterns and abstractions, leading to the conclusion that software designed for AI should align with these established mental models [2][4] - A good mental model is stable and extensible, allowing for new implementations without disrupting existing structures, as seen in file systems like Linux VFS [5] Group 2: Importance of Software Ecosystem - The software ecosystem's significance is nuanced; while syntax and protocol differences may seem trivial to AI agents, they still reflect stable mental models that are crucial for effective training [6][7] - The real importance lies in whether the software's underlying model is robust and well-validated, as this determines the agent's ability to adapt and utilize the software effectively [7] Group 3: Interface Design for AI Agents - Effective software interfaces for AI agents should meet three criteria: they must be describable in natural language, solidified in symbolic logic, and deliver deterministic results [8][9] - Natural language is suitable for expressing intent, and AI models have become adept at interpreting ambiguous language, making it a viable interface for agents [11][12] - A successful system should minimize ambiguity in its internal representation, allowing for clear execution of tasks and enabling agents to operate efficiently [14][15] Group 4: AI Infrastructure Characteristics - AI agents produce workloads that are inherently disposable, emphasizing the need for infrastructure that allows for rapid creation and abandonment of resources without significant overhead [22][21] - The emergence of AI agents has lowered the barrier to writing code, making previously unfeasible demands viable, thus expanding the range of user needs that can be addressed [24][36] - Infrastructure must support extreme resource sharing while providing a sense of independence to agents, allowing them to experiment freely without impacting others [28][30] Group 5: Changes in Business Models - The advent of AI agents has made previously uneconomical business models feasible, as the cost of fulfilling long-tail demands has significantly decreased [36] - Successful AI agent companies should focus on transforming single-use computations into scalable online services, thereby reducing marginal costs and enhancing sustainability [37][39] - The shift in user demographics, driven by AI agents, necessitates a reevaluation of traditional cloud service models, emphasizing the need for adaptability and efficiency in service delivery [38]