Rethink the terminal: A portal to AI

AI CLIs have fundamentally changed what I use a terminal for. I now use it for almost all my work: coding, planning, review and writing, from starting a task to finishing it and revisiting it later.

My focus is on understanding and steering agents’ work rather than running commands. That calls for a terminal designed around this collaboration, both inside the window and across the desktop.

I built AgentTerm as an experiment and have been using it since January. These notes share some of the thinking behind its design.

Rethinking the terminal

Why I give each lead agent its own OS window and a unique Dock tile or taskbar button

Why I run AgentTerm from source

My agent changes the source; the next window runs it. No installer dialog or launch command.

Why documents, code and images open right in the agent’s terminal window

The agent writes; we collaborate on the rendered content.

Plain text as an interface to build on

Building on the terminal

Beyond the terminal itself, I have been exploring how its new capabilities can improve the way I work with agents.

Curated code review: explanations grounded in code

I removed a check so the agent could focus on what needs my judgment.

A hub for agents across machines, with phone and voice access

The agent interprets my spoken instructions in the context of our work.

AgentTerm has other features and design choices not covered above, such as writing with coding agents, commenting on their terminal output or copying it without reformatting, connecting to an IDE, running long jobs, coordinating a shared workspace, and setting up and invoking guides across AI CLIs using a namespace convention based on folder paths.

Check out the AgentTerm repo, try it for yourself, and build on it.