Vercel AI SDK Generative UI Design

Honesty about what it can do and visibility into what it is doing. Interfaces that overstate capability get abandoned the first time they fail, and interfaces that go silent during a long tool call get treated as broken even when they are working. Showing intermediate progress, citing sources, making cancellation always available, and failing clearly rather than confidently do more for adoption than any amount of visual polish.

No. Streaming is right for prose a user reads as it arrives, but it is actively worse for structured output such as a table or a form, where a half-rendered result is unreadable and slightly unnerving. The useful pattern is streaming prose, progress indication for structured work, and a clear completion state so the user knows the system has finished rather than stalled.

By showing the steps. A user waiting thirty seconds with a spinner assumes it is broken; a user watching the agent report which step it is on will wait considerably longer and trust the result more. Cancellation must stay available throughout, and partial results should survive a cancellation rather than being discarded.

Yes, and streaming content makes it genuinely harder than a static page, which is why it needs deliberate design rather than a retrofit. Live regions have to announce new content without interrupting constantly, focus must not jump as tokens arrive, and every interaction has to work from the keyboard. We build this in rather than adding it after an audit finds it missing.