📄 中文摘要
回顾十八个月前的AI发展历程，那个时候的上下文窗口较小，'提示工程'成为热门技能，而MCP尚未出现，AI代理自主操作外部服务的概念仍然是理论。作者在日本大阪构建医疗AI产品时遇到的问题，成为了2026年一系列变化的种子。这一系列的变化揭示了AI代理未来的发展方向，尤其是在技术不断进步的背景下，如何更好地利用AI技术来解决实际问题。
并非所有事物都需要MCP，第二部分：2026年相变——三条独立道路通向同一结论
📄 English Summary
Not Everything Needs MCP, Part 2: The 2026 Phase Transition — When Three Independent Roads Led to the Same Conclusion
Reflecting on the AI development journey from eighteen months ago, the context windows were small, 'prompt engineering' was the sought-after skill, and MCP had not yet emerged, making the concept of AI agents autonomously operating external services largely theoretical. The author faced a problem while building a medical AI product in Osaka, Japan, which contained the seeds of the changes that would unfold in 2026. This series of changes reveals the future direction of AI agents, particularly in how to better leverage AI technology to solve real-world problems as advancements continue.