What makes an agent chat feel good?

Agent responses often include much more data than just their reply, so a chat has to decide what to emphasize. Getting that emphasis right separates a readable chat from a wall of output, and it means three things: the reader knows what the agent is doing, their attention stays on what matters, and each piece arrives where and when they expect it.

Progress belongs on one line

A transcript feels rushed when many things move at once, even if little text arrives. Anything that appears pulls the eye,¹1Yantis and Jonides, Abrupt visual onsets and selective attention: Evidence from visual search, Journal of Experimental Psychology: Human Perception and Performance, 1984. so each one costs attention.

While a turn runs, one status line should say what the agent is doing, updating in place from its tool calls, thinking and short remarks. Knowing what’s happening makes the wait feel shorter.²2David Maister, The Psychology of Waiting Lines, 1985.

But a line that updates with every step is always changing, so each phrase needs time to be read. Subtitles have the same constraint: in an eye-tracking study, most viewers kept up with them at 20 characters a second, the fastest speed it tested,³3Szarkowska and Gerber-Morón, Viewers can keep up with fast subtitles: Evidence from eye movements, PLOS ONE, 2018. and Netflix uses that as its limit for adult programs.⁴4Netflix, English (USA) Timed Text Style Guide, reading speed limits. At that rate, a phrase like “Reading src/auth/session.ts” takes about a second and a half to read, and it should stay up at least that long. If more steps happen in that time, the line should skip to the latest. A phrase that changes before anyone can read it is only a distraction.

Everything except the agent’s replies should collapse

During a turn, the reader should see what the agent is doing and anything it wrote for them. Short remarks between steps belong on the status line. Longer messages, or ones with a list, heading or code, are generally written for the reader, so they should stream into the transcript and stay there.

When the turn ends, the reader should see the reply, with everything the agent did folded into one closed line above it. The reader gets what they need now, and the rest stays one click away, a pattern known as progressive disclosure.⁵5Jakob Nielsen, Progressive Disclosure, Nielsen Norman Group, 2006.

Text should reveal steadily instead of in bursts

Agents don’t send text at an even pace. It comes in uneven chunks, and some agents only send a message once it’s finished. Uneven chunks look like stutter even when the average speed is fine. The fix is to hold text back and reveal it at a smooth pace instead of as it arrives.

The more text is waiting, the faster it should reveal, aiming to clear the backlog within a set time, the catch-up target. A minimum rate, the floor, keeps the last few characters from crawling, and a maximum, the ceiling, keeps a big chunk from arriving as a jump. The cost is that the screen trails the agent, and these settings decide by how much. The table shows it for a 1,000-character message that arrives all at once, with rates in characters per second (cps).

With the slow settings, a hosted agent feels as slow as a model running on your laptop. Pacing only has to even out bursts. Unlike a status phrase, a reply stays on screen, so it doesn’t have to reveal at reading speed. Once it keeps the reader waiting after the agent has finished, it’s holding back text they could already be reading.

It also matters how far the text moves on each frame. Small, even steps of a character or two read as smooth motion. Revealing by word makes the steps uneven, since words vary in length, and the eye notices that for the same reason it notices bursts. In one study of scrolling text, comprehension dropped when it advanced in the largest steps, and readers preferred it moving smoothly.⁶6Kang and Muter, Reading dynamically displayed text, Behaviour & Information Technology, 1989.

A reader moves from one line to the next while new text keeps shifting the page, and that only works when the shifting is steady. Bursts make it irregular, so the reader keeps losing their place.

The transcript should fill top to bottom

Pacing only works if it covers everything. A message that’s still revealing can be overtaken by whatever the agent sends next, so a short reply or the fold appears underneath it in full and pulls the eye down and back. The whole transcript should be paced as one queue, so it always fills from the top down.

Auto-scroll should follow the reader

The view should stay at the bottom while the reader is there, and stop following the moment they scroll up to reread something. Only the reader’s own scroll should stop it. New content also pushes the bottom away, so distance from the bottom can’t tell whether they moved. Coming back to the bottom should start it again.

Each of these changes is small, but over a whole turn they add up:

None of the techniques shown here are about the agent itself. They’re about the reader, and not wasting their attention.