DEV Community

Cover image for Turn 300 shouldn't feel slower than turn 3
Manuel Lopez
Manuel Lopez

Posted on AI-assisted

Turn 300 shouldn't feel slower than turn 3

You know the moment. The agent has been at it all afternoon: a few hundred turns, builds, test runs, long diffs. Now you need the window more than ever. You want to scroll back and see what it changed, read the last diff, answer the permission prompt it's waiting on, and send it the next instruction.

And that's when the interface starts to drag. Keys land a beat late. Switching to another agent's tab makes you wait. The session you care most about, the long one with all the context, is the one that feels heaviest.

I'm Manu, and I build Lichen, a free Windows app that runs your Claude Code and Codex sessions side by side, each in its own tab. I keep twenty-odd sessions open all day, and many of them run to hundreds of turns. Lichen had this problem too, so we went after it until the length of the session stopped mattering.

The number

In our lab benchmark (jsdom, not the live app), a 300-turn conversation re-renders in about 12-13 ms per tool event (median; p90 under 25 ms), down from 1.8 s before the fix.

It no longer grows as the conversation gets longer. Turn 50, turn 150 and turn 300 all land in the same range, so the session you've been working in all day costs the window about the same as one you opened ten minutes ago.

A tool event is what reaches the window every time the agent runs something and gets a result back, and in a working session they come in bursts. At 1.8 s each, those were seconds where the window belonged to the agent and not to you. Now they're milliseconds.

What the number measures

  • It comes from our lab benchmark (jsdom, not the live app) with a 300-turn conversation, run on the Lichen 2.0.2 release. It's not a stopwatch on the installed app.
  • It measures Lichen's own interface, not the model or the CLI. The model doesn't answer any sooner: a long session still sends a long context, and that wait is the same as in the terminal. What changes is everything around that wait.

What you notice

The difference isn't on turn 3, when every tool is fast. It shows up at turn 300:

  • You type and the text is there. You can write the next instruction while the agent is still working, without the keyboard falling behind.
  • You switch tabs and the other session is there. Going from the agent that's building to the one that's waiting for you doesn't come with a pause.
  • You scroll back through hours of work to find what changed, and the window keeps up.

That's what makes twenty-odd sessions workable in one window. Each tab tells you whether its agent is working, waiting for you, or finished while you weren't looking. You can only act on that if getting to the tab is instant.

How it stays fast

Mostly by not redoing work that didn't change. In a long session, every tool event used to make the interface redo the whole conversation, so the cost of one event grew with every turn before it. Now an event only updates what changed, and turn 300 costs about what turn 50 does.

What it doesn't change

  • The model's speed. Covered above: Lichen makes the window faster, not the answers.
  • How you use Claude. Lichen runs on top of the Claude Code CLI you already have, no -p, no SDK. Codex sessions use the Codex CLI you already have.
  • Your plan usage. Does a session in Lichen use your plan differently from the same session in a terminal? We haven't measured it yet; we will, and we'll publish the numbers.

Try it on your longest session

Lichen is free, runs on Windows, and there's no Lichen sign-up. Install it, open your longest Claude Code or Codex session (or pick up an old Claude Code conversation from the history), and see if it keeps up with you.

Download Lichen for Windows, free

How long do your agent sessions get before you start a new one just to make the window usable again?

Top comments (0)