Try the redesigned human-in-the-loop interaction flow — execution plan review, mid-task modification, and sign-in handoff.
Built directly in code with Claude Code. Every interaction — the typing animations, the exact moment the split view slides in — was implemented and iterated by hand, because interaction design only comes alive when you can feel the timing. Building live also surfaced ideas that wireframes never would. You can trace how some of these components evolved on the Design Evolution page.
What's in the Demo
Take your time with it. The typing animations are intentional and paced to feel like the agent is actually thinking.
Before the agent runs, it presents a step-by-step execution plan for your review. Approve it as-is, or modify individual steps before anything begins. This is the primary human-in-the-loop checkpoint.
Mid-task, the agent encounters a step requiring Google authentication. Rather than failing silently, it pauses, explains why it needs your help, and hands control back cleanly.
