aywu

I wrote about building a better terminal for AI, rather than an IDE or a desktop app. I like the terminal's empty starting point and consistent text surface: they help me focus, take in more material and form my own conclusions.

How do you see the tradeoffs between terminals, IDEs and desktop apps as all three evolve around AI? And what do you think of AgentTerm's approach?

I did think about adding a messaging channel through the stream hub, which is the relay behind the phone (web) viewer. But I did not built it because lack of strong needs. I sometimes copy or rephrase messages between agents, for example, if I hit an AgentTerm bug in a session, I take it to a separate session to avoid context getting polluted.

What kinds of things do you have Codex and Claude talking to each other? I am interested in real examples to understand better.

My agent often kicks off CI and waits for the result. It ends its turn right after the kick-off, while CI is still running, so it can respond to me immediately. Which state does that cell get: working, needs you, or done? It does not need me at that point, but it is not done either.