The Faster AI Runs, The Easier It Is for Us to Get Lost in Front of the Screen?

In 2026, AI coding agents have crossed the million-token context window — they can search entire repos, understand dependencies, and carry out tasks autonomously. The models have improved significantly, taking over a vast amount of our technical workload. However, human working memory is still limited to 7 ± 2 items. While AI processes millions of tokens seamlessly, developers facing dozens of open editor tabs still frequently fall into the trap of "Who am I? Where am I? What was my next step?"

This talk follows the story of two niche, open-source VS Code extensions I built in my free time. The first, Quick Prompt, addresses the asynchronous nature of human-AI collaboration ("while the AI runs task N, the human has already thought of N+1") by providing an internal scratchpad to offload memory and maintain cognitive continuity. The second, VirtualTabs, is a virtual file directory for organizing editor groups. After implementing integration features like AI context exporting and an MCP (Model Context Protocol) server, the few issues I received were surprisingly focused on fundamental UI needs—highlighting the active file, drag-and-drop sorting, and window layout management—features that help humans maintain spatial awareness.

This session will primarily focus on practical development experiences, sharing the technical implementations from these projects as well as the philosophical reflections inspired by user feedback: In an era where models can solve most technical problems on their own, should our tooling efforts focus on the bottlenecks of AI, or the cognitive bottlenecks of humans? If you are interested in developer tools or human-AI collaboration experiences, you are welcome to join this session.