Skip to main content
Nobody writes the agent’s code here. You say what you want, the agent does the work — including editing its own harness and publishing a version of it — and you check what came back. These five pages are that skill, worked through five times on a real account: what was said, what the agent did, what showed up on the screen, and how it was checked afterwards. Where a page shows Python, it is the code the agent wrote when it was asked, kept so you can see what happened. It is evidence, not homework.

Your first real result

Ask for a job that ends in a file. Then find the file from a second chat, take it off the machine, and look at it yourself.

A tool of your own

Ask for an ability your agent does not have. It writes the tool, tests it, wires it in and publishes a version of itself.

A job that runs every day

Ask for a briefing every morning. The agent writes the program, starts it, and it keeps running after the chat ends.

A Telegram bot

The same shape with a different trigger — and the one failure worth knowing about, as it actually happened.

Drive a run from your own code

The one page for a reader holding an API key: create a run over HTTP, stream its events, read its outputs.

​
The shape they all have

Every page follows the same four beats, because that is what using this product is:
  1. You ask — in your own words, for the outcome you want, not for the code.
  2. It does the work — and you watch the tool calls, which are the agent writing and running code on its machine.
  3. You check — the file is there, the version number went up, the process is running, the second chat sees it. Checking is a message, not a console.
  4. You undo it if you were wrong — also a message. Every page ends with one.

​
What you need

An account and the machine that came with it. Signing up gives you both, plus a harness copied from the assistant starter. There is nothing to install. A stopped machine is fine. Sending a message in a chat that runs on it starts it. You never have to press Start first. A paid plan for one of the five. Drive a run from your own code needs an API key, and minting one is refused on a free account. The other four work on any plan. A Telegram account for the bot — the token comes from @BotFather, and that is one of the few steps here your agent cannot do for you. Patience, and some of your plan’s usage window. Asking for a file back is a minute. Asking your agent to rewrite its own harness — read the tree, write the code, test it, publish it — is a run that thinks for the better part of an hour and costs accordingly. Two of the pages here are that kind. A plan meters usage over a rolling five-hour window, and a day of this can reach the end of it:
That is a refused run, not a lost one — what was already published stays published. Usage shows where you are in the window.

​
What cannot be done by asking

Worth knowing before you start, because these are the places where you have to get up and do something:
  • Anything that needs a human identity somewhere else. Registering a Telegram bot, approving an OAuth consent screen, accepting terms in somebody else’s product. The agent can use the credential once it exists; it cannot be you.
  • Minting an API key. It is done with your browser session, on a paid plan. See API keys.
  • Paying. Plans, top-ups and limits are yours. See Billing.
Everything else on these pages — writing files, writing tools, publishing versions, starting long-running programs, checking any of it — is a sentence in a chat.

​
Reading the machine yourself

You do not have to take the agent’s word for it. In a chat, the inspector panel on the right offers Open in VS Code and an SSH command for the machine this chat runs on, good for an hour. That is how you read what the agent wrote — the harness, the files, the logs — without asking it anything. See The chat.

​
The order they are in

They stand alone, and the first one is the one to start with — it is the shortest and it teaches the habit the rest depend on: the work lands on a machine, and you check the machine. After that, A tool of your own and A job that runs every day are the same act seen twice — the agent editing its own harness and publishing a version of itself — once for an ability, once for something that runs without you. Read either first. The Telegram page is the job page with a different trigger, plus the failure that catches everybody.
These pages are not a reference. When you want the whole mechanism rather than one worked case, the pages that have it are Concepts for the words, Changing what your agent is for the harness, and API for the endpoints. Those pages are written for your agent to read as much as for you — hand it the URL and it will.