From Prompt to Deployed Site on a Cloud Computer

The difference between "here is code" and "here is your site running" is a computer. This is the end-to-end run: one prompt, one agent, one cloud machine, a URL at the end.

The prompt

Spin up a computer, scaffold a static site for a bakery (menu, hours, contact form stub), run it, fix anything broken, and give me a public URL to review.

That is genuinely all of it. What happens next is the interesting part.

What the agent does

Provisions a cloud computer: a real machine with a shell, not a toy sandbox.
Scaffolds and installs: writes the files, runs the package install, sees the same errors you would.
Runs and fixes: starts the dev server, reads the failure, edits, retries. The write-run-fix loop is where agents with computers stop being autocomplete.
Exposes a tunnel: a public URL to the running server, so "review it" means clicking, not cloning.

You watch the run trace throughout: every command, every edit, every retry is inspectable, and the autonomy dial decides how much of it pauses for your confirmation.

Where this pattern earns its keep

The bakery site is the demo; the pattern is general:

Reproduce a bug on a clean machine with the exact dependency set.
Try a library for real (install, run its examples) before adopting it.
Run a data job that needs real dependencies and hours of wall-clock, without your laptop staying awake (code migrations scale this up).

Good to know

Computers bill by usage with the price visible; stop them when done (agents can do that too).
Your own machines pair the same way when the job should run on hardware you own.
Budgets cap the run's spend; per-computer permissions cap its reach.

Run the bakery prompt as your first computer task: it exercises the whole loop in fifteen minutes and leaves you a URL to show for it.