







There's long been a "100ms threshold" for making software feel instantaneous.
I sometimes wonder if there's a similar threshold to be discovered for the UX of AI, but for optimal message lengths.
If you're using AI and you're in the inner loop (talking back and forth, not delegating and walking away), there's clearly a limit to how many "operations" one can absorb per turn.
Almost all models skew to too many words and too many changes per turn.
This threshold is probably much smaller than one would think, perhaps the length of a short WSJ op-ed.
This max-ops threshold would apply to operations like code edits and also to the narrative around it.
The ideal inner-loop flow state would combine near-instant feedback with this max-ops-per-turn threshold to create a continuous flow state with the agent. On the product side, I suspect a UX like this would correlate with measures of stickiness and engagement.
Obviously, this UX would not be optimal for long-running tasks. You could imagine opting in to these different modes: one for delegation, and one for an interactive flow state that adheres to this threshold.


















