Lose the harness. Let it cook.

A persistent computer where any model can use real tools, run real software, and produce real work.

premise

AI is remarkably capable. We keep surrounding it with abstractions designed for weaker models.

No restrictive tool registry
No predetermined workflows
No skill ceremony
No vendor-owned environment
No guessing what the agent might need

Give it real tools and get out of the way.

The abstractions go on the human side. The agent gets a computer. Where this goes.

what you get

One Linux box, yours, that keeps everything. Ubuntu with the tools an engineer expects, a user with sudo, and a home directory that survives.

/home/suped/
  workspace/    where shells open
  projects/
  downloads/
  .config/

bash · python · node · git · curl · wget · jq · sqlite · ffmpeg · ripgrep · unzip · uv · Chromium via Playwright. Need more? Install it. It stays.

No special filesystem abstraction. No suped-specific anything inside the box. Linux is Linux.

how it works

Run the command. The first run builds a local image and creates a persistent home. Every run after that is instant.
Work in it, or hand it over. Open a shell yourself, or point an agent at it with suped exec. Install tools, clone repos, log in to services.
Come back to it. Swap agents, swap machines, swap models. The computer is still there with everything you set up.

the prompt

This is the entire system prompt. suped prompt prints it.

You are operating a persistent Linux computer on behalf of the user.

You have access to the shell, filesystem, installed applications,
and explicitly connected services.

Use the computer to accomplish the user's objective. Inspect the
environment, install dependencies when appropriate, write scripts,
use APIs and CLIs, and preserve useful work in the filesystem.

Ask the user only when you need information, authentication, or
approval for a consequential action.

That is practically the whole philosophy. Read the rest.