Give Your Agent Its Own Computer

This repo goes with my September 23, 2026 O'Reilly session, Give Your Agent Its Own Computer, part of Zero to Agent in 30.

In the session, I give an agent a computer of its own in the cloud. It can use that computer through the shell or through the screen, and we watch it do a real task each way.

Each demo is one script that reads top to bottom. The prompt, the model, and the settings are at the top of the file.

• Python 3.12
• uv
• An E2B API key. The free plan is enough.
• An OpenAI API key with access to gpt-5.6-sol and gpt-5.6-terra. These calls cost money. See Costs.
• A web browser, to watch the agent's desktop in demos 2 and 3.

git clone https://github.com/sajal2692/zero-to-agent-own-computer.git
cd zero-to-agent-own-computer
cp .env.example .env
uv sync

Open .env and add your two keys. Git ignores this file.

uv run python demos/01_shell_data_analysis/agent.py

The agent gets this task:

I'm helping my niece choose a college major. The FiveThirtyEight college majors data is in this GitHub folder. Download the CSV files and work out which majors pay best and which pay worst for recent graduates, and where a graduate degree makes the biggest difference. Check whether the high earners also have low unemployment. Write me a one-page PDF report with a couple of charts.

Watch it install its own packages, run the analysis, and check its finished page before it stops. The script then brings report.pdf back to your machine and deletes the sandbox.

uv run python demos/02_desktop_agents_sdk/agent.py

The script prints a link. Open it in a browser to watch the agent's desktop, then press Enter in the terminal to start. The agent gets this task:

I'm setting up a home office in a small room. Look on IKEA Canada and find me the three best desks under 250 Canadian dollars, no wider than 120 centimetres, rated four stars or better. Write me a short report with the three desks, their price, size, and rating, and tell me which one you'd pick and why.

It starts from an empty desktop, opens the browser itself, and works through the site one screenshot at a time. It ends with a short written report.

uv run python demos/03_desktop_plain_loop/agent.py

Same task and same desktop as demo 2. This script has no agent framework. It calls OpenAI's computer use tool directly, so you can read the loop line by line: observe, choose, act, check.

The desktop scripts take a task as an argument:

uv run python demos/02_desktop_agents_sdk/agent.py "Find the opening hours of the Art Gallery of Ontario this Saturday."

For demo 1, edit TASK at the top of the script.

In every demo the agent loop runs on your machine, and the work happens on the agent's computer. Each log line says which:

• LOCAL is your machine: creating and deleting the computer, the task, the model's thinking, the final report, and the cost summary.
• SANDBOX is the agent's computer: each command and file operation in demo 1, and each click, scroll, and keypress in demos 2 and 3.

Each run saves a timestamped folder under that demo's output/ directory:

• report.pdf for demo 1, or report.md and final_screen.png for demos 2 and 3
• run.json, with the time taken, token counts, and the cost of the run

Git ignores the output/ folders. The sandbox is deleted at the end of every run, including after an error or Ctrl+C.

These are typical figures from my own runs in September 2026. Yours will vary.

E2B usage comes out of the free plan's one-time credit. A desktop sandbox costs about half a dollar an hour, and a run uses a few cents. The free plan limits a sandbox to one hour.

The price constants at the top of each script feed the cost summary. Check them against the current price lists.

Nothing in this repo is synthetic. Demo 1 downloads the college majors data from FiveThirtyEight, licensed CC BY 4.0. It comes from the American Community Survey for 2010 to 2012, so treat the numbers as an example and not as current salary information.