Skip to main contentSkip to navigation
All posts

Notes on working with AI

An always-on agent needs exactly one job

Agents that keep working with your laptop closed are here. What makes one useful is simple: one job, a manual run first, and approval on anything final.

4 min read

The laptop is closed and the agent is still working

A new kind of AI product has arrived, and xAI's Grok Bot is the clearest example right now. It isn't a chat window that waits for you. Each Bot has a standing role, a cloud computer with a browser, a terminal and files, connections to your apps, and a schedule. You can tell it to check the inbox every morning and draft replies, and it does that at 8:00 whether your laptop is open or not.

That changes the question people ask. With a chatbot you ask "what prompt do I write?" With an agent that works while you sleep, the question becomes "what job do I give it, and what is it allowed to do without me?" Most of the setup that matters is the answer to that second question.

How it works under the hood

Four pieces, and once you see them the product stops looking like magic.

A computer. Every account gets one persistent cloud machine. When there is no work it goes to sleep, and it wakes when a Bot needs it. You can watch the screen live and see the cursor move as the Bot opens a site and reads it.

Connections. Gmail, Slack, Notion and others connect once for the whole account, and anything with an MCP server can be plugged in too. Logins to ordinary websites happen through a secure form, so the password goes to the site and not into the conversation.

Skills. A skill is how to do something: a reusable set of instructions, shared by all your Bots. "This is how I want the weekly report to look."

Routines. A routine is when to do it. It belongs to one Bot and runs on a schedule or on an event such as an email or a Slack message. xAI's docs set a few hard limits: runs at least five minutes apart and up to 50 routines per Bot.

Skill is the how, routine is the when. Keep those apart in your head and the whole system becomes easy to reason about.

The detail almost everyone misses

All your Bots share the same computer. Each Bot has its own screen, but the browser sessions, files and saved logins are common to all of them. xAI's own documentation puts it plainly: the screens are separate work surfaces, not separate security boundaries.

In practice, if you log in to your bank on that browser for a finance Bot, your marketing Bot can reach it too. So the rule is simple. Don't put a login or a file on that machine unless you're fine with every Bot on the account using it.

Two more things from the docs that people learn the hard way. A test run of a routine does the real work: if the routine sends an email, the test sends the email. And the Bot asks for approval before actions that can't be undone, such as sending, publishing, buying or deleting. You can approve once, approve always, or refuse. Start with "once."

One Bot, one job

The setup that works is narrow. Not "a general assistant for my business," but one Bot that handles customer inquiries, and another that tracks cash flow. A narrow role gives the Bot clear rules, and it gives you a clear way to check whether it did the job.

The order I recommend:

  1. Run the task once by hand in a chat with the Bot, on safe input.
  2. Check the result line by line.
  3. Turn it into a skill once it comes out right.
  4. Only then put the skill on a routine.

Skipping straight to step 4 is how people end up with an agent that confidently sends the wrong thing every morning.

Where to start this week

Pick the task that eats an hour of your week and has a clear finished state. Some good first jobs: a morning summary of new emails with draft replies that wait for you, a weekly check of what competitors published, or a monthly list of recurring charges on a business card that nobody remembers signing up for.

If you want to try Grok Bot specifically, the cheapest plan that includes it is currently listed at about $30 a month, and the official docs are the best place to read the limits before you pay. Whatever tool you pick, the discipline is the same: one job, a manual run first, and a human on anything that can't be undone.

Claude training for teams

Want your team working this way?

This is the material we teach in workshops, at your office or online. Tell us what you're trying to solve, and we'll come back with a plan.