"it runs on your computer"

mech suits and tuis

Karpathy’s 2025 LLM year-in-review created some confusion around his language about Claude Code. A portion of the audience couldn’t grasp, or didn’t like, that the framing was “Claude Code runs on your computer.”

I don’t think we should push back on this idea. The agent is on your computer.

And this is exactly how we should think of agents moving forward. And not like any type of web harness or agent platform terminology that has been made irrelevant since the dawn of Claude Code. These are what agents look like. Not “deep agents”, not “computer use agents”, these are agents.

The Mech Suit

Think of an agent like a mech suit plugged into a brain in a jar far away. The intelligence is remote, but the suit is here. With you. On your machine.

It can operate on your entire operating system. Claude Code’s Bash() tool is a shell. Given a shell, it can do anything on the OS that is programmatically possible. It can move files, read directories, execute commands, spawn processes. It can traverse your file system. It can reach across networks on your behalf.

It is here. It is a thing on your computer.

That’s why it’s different. That’s why Anthropic acquired Bun — they’re going after the agentic runtime. What I foresee is an operating system that we build apps on. Agent Skills are a first look at that.

On TUIs

There’s a vocal portion of Twitter that hates Textual User Interfaces. I think they misunderstand the UX here.

What we’re doing these days is orchestrating agents, sub-agents, background tasks, tool calls. We’re managing concurrent operations, monitoring outputs, directing workflows.

Throughout history, there hasn’t been a better interface for orchestrating systems than CLIs and TUIs. There’s a reason we don’t have a graphical user interface for kubectl.

I do foresee graphical UIs developing, but not anytime soon. I imagine these UIs looking and feeling like games for operating factories of sorts. But we’re far away from there.