Most people trying to win more customers are still the whole company.
Find the names. Write the letters. Make the pictures. Answer the ones who wrote back. Count the week. Every one of them waits for the one before it to finish.
Nothing gets handed off. Everything waits its turn, because one head and one pair of hands can only be in one place, and by the time the day is gone the list is still sitting there.
Most of them already know the chasing and the counting never needed them in the room. They do it anyway, because there is nobody else.
This is how that one-person line becomes a company. You give the jobs names, put them on one shared computer, and the first hour of the day is already moving when you sit down.
By Wednesday of the first week the roster looked like this.
- Mara: finds the people who would buy. She comes back with a list, not a summary.
- Cole: writes to them. One letter per name, in their language, ready to send.
- Rina: makes the pictures. The landing page, the post, the banner that runs next to the letter.
- Vince: answers the inbox. The ones who reply get a human answer the same hour.
- Owen: counts the week. Who paid, who went quiet, and what it was worth.
Five names on one computer, and it has room for 50.
Read it to the end. The last block is the one move that keeps the whole company alive after that computer updates itself: you are already root on it, and only one folder survives the wipe. I will show you exactly what to put there.
Before we get into it, follow me on X and join my Telegram channel where I post more AI content every day. Both are free.
X - https://x.com/Argona0x
Telegram - https://t.me/+r0clI4-MMC03ZjAy
What Grok Bot actually is
The window looks like a messenger, with names down the left side and bubbles down the right. That is why the first day usually goes the way every other chat has gone: you type a question, you read an answer, you type again.
Behind the bubbles, each name is a teammate with a job, a conversation of its own, and a working memory that does not get wiped between tasks.
Every one of them also gets a computer with a browser, a file system and a terminal, and that computer runs in a data center, not on your desk.
Close your laptop halfway through a task and the work keeps going. The FAQ asks *"Does Grok Bot keep working when my laptop is closed?"* and begins the answer with one word: *"Yes."*
One account gets one computer, and every employee on it works on its own screen.
Several of them run at the same moment, and because the files, the browser sessions and the logins all live on that one shared computer, whatever one of them finishes and saves is already sitting there for the next one to pick up.
Inside the company that built it
The company that built it works this way in-house, and says so in its own launch post: people there often keep several bots going at once, with one on top to manage the rest.
A chief of staff, then a specialist for each lane: inbox management, expenses, recruiting, bug fixes, operations.
Matt Palmer does developer experience inside that company and had it before launch. Every personal agent until now, he wrote on 11 August, left him *"configuring, provisioning, delegating, and triaging instead of doing useful things."*
He also had one of them grind Wii Baseball on an emulator overnight. It learned the controls over the course of a few games and won.
All of it runs on the same model you already pay for.
*Somebody just got hired to use it.*
Step 1 · Hire your first coworker
You have one chat window open and nobody in it. The account holds 50 Bots and group chats combined, and yours is using one.
Fletcher Richman spent launch day arguing against using it. He ran product for Confluence and now builds type.com:
A vanity metric, in his words. He added the reason the next morning: *"now they are having to manage a ton of teammates? most people are terrible managers."*
@MelvinNerdster does not write code and runs three of them. A chief of staff, a note-taker called Betty, and one he calls Elvira, the fly on the wall that checks every output before the next step runs.
Handing over one part of his workflow gives him back 2 to 4 hours a day.
What deserves its own employee
The first hire is one job. The manual gives five tests, and the job has to clear all five to deserve its own employee:
- goal: one area of ownership that belongs to nobody else.
- sources: its own set of tools, apps and files.
- working style: a way of working that differs from the rest of the roster.
- approval boundary: a named line where it stops and asks you.
- schedule: the job repeats on a clock.
Talent Scout, Expense Manager and Bug Reproduction clear all five. General Helper clears none: the docs say it gives the bot less guidance and makes its saved context harder to reuse.
Hiring is four moves:
- new chat: *Cmd/Ctrl+N*, or *New* in the sidebar.
- create new agent: choose it there, and a bot called *New Agent* appears.
- *Bot actions* → *Edit Profile*: name, title, description, avatar.
- first message: a real job, never a greeting.
The names people land on are Paul, B1, Stitchy, Dex. Paul Lipsky asked his to pick its own and it came back Atlas.
The description is the hire
The description is the field people waste. It holds the rules that stay true after you have forgotten writing them.
Today's work goes in the message.
Both role cards in the manual end on a limit. One closes with *"Never contact a customer or change an account without approval,"* the other with *"Never change production settings."*
Mara is the prospector, and hers is built to that shape:
Four sentences, two of them limits. You can check her work against the description without reading the whole run.
Her first message is today's work, in five fields. Leave one field out and she stops to ask you about it before she starts:
- outcome: what has to be finished.
- sources: which apps, sites and files to work from.
- constraints: what to avoid, and what to ask about first.
- deliverable: what to hand back.
- review point: where she stops for you.
*The description you just wrote is also everything the next hire will ever know about Mara.*
Step 2 · Let them hand work to each other
You hire the second one and the courier job becomes yours. Mara turns up a name that belongs to Cole, so you copy it across while Rina waits on her brief.
The roster fills up and the work still moves at the speed you can carry it.
@jaehwa pulled apart Debbie O'Brien's chief-of-staff output on 14 August and quoted that bot's own diagnosis. In translation:
bots that have no role description cannot hand work to each other
A blank description takes that bot off the chart. A task that belongs to somebody else moves on what the other descriptions say, and nothing else.
*A bot's description is the org chart.*
The handoff itself is an asynchronous message. A bot sends it, the receiving bot wakes, handles the request and replies later.
The whole exchange stays visible in the conversation where you can read it.
A group chat takes 2 to 6 bots.
Write normally and they work out between themselves who answers, with *@* there for the times you want to name one yourself, and if you ask for a single owner at each stage the parallel handoffs stop duplicating work.
The opening line the docs give a new group, verbatim:
Three roles, one deliverable and a stop rule, with your names in place of these.
Debbie O'Brien's prompt goes the other way and makes the bot write the chart. Elon Musk reposted it as *"How to use @bot"*:
It reads your roster and returns the descriptions you never wrote. A day later she posted what came back:
She asked the chief of staff for one thing. It messaged the other bots, they messaged back, and a pull request arrived for her to review.
Copy the one that already works
Cole works one market and you want a second. Duplicate him:
- carried over: profile, settings, enabled skills, routines and avatar
- not carried over: conversation history, learned memory and chat attachments
- naming: the copy arrives as *<name> copy*, so rename it before it takes work
The copy lands on the chart with the same job and no history.
Whatever they figure out on the way dies with the conversation. Tomorrow morning every one of them starts the job from scratch.
Step 3 · Teach the job once, then schedule it
Cole's letters this morning had the same two mistakes you fixed on Monday. He kept none of it, so the method still lives in your head.
The docs answer that with an order. Five moves, and the schedule is last:
- run one real task: a live job with a scope that is safe to get wrong
- correct the result: keep going until the output is something you would review
- save the process: the run that worked becomes a named skill
- test it on a second input: a different account, a different file, the same steps
- create the routine: only once retries and failure cases are written down
The schedule only works because the four moves above it already ran.
Rina's job is 15 clicks in a browser, so *Teach a task* records it.
The recording runs from the computer view for up to 10 minutes, takes no microphone audio, and hands back the method as a draft you can edit.
Draft is the docs' own word for it. A single recording cannot show what to do when a source goes missing or a price needs approval.
Say this to it, word for word:
Sources, rules, output shape and the approval line, fixed in one pass.
*A skill outlives the conversation that taught it.*
A skill still waits to be called. The routine is the call, and one sentence sets the time, the source and the stop rule:
That last clause keeps yesterday's numbers out of Monday.
One bot holds up to 50 routines, and the 20 most recent runs of each stay on record.
Let the bot write the routine
Describe the job in conversation and the bot writes the routine itself. Ray Fernando, in a 74-minute build stream on 11 August:
you can just say, can you create a routine, just explain it in here. And then they can create the routine for you on their behalf
His weigh-in check runs every morning and reports on Sunday, and he never wrote it.
The routine panel is where the tools plug in as well: Slack, Git, Linear, Sentry, PagerDuty.
Every one of them still hands its work back into a chat window, because not one of them has a login to anything.
Step 4 · Hand them the real accounts
Vince has the refund reply written and it is sitting in a chat window. You paste it into the mail client yourself, the way you do at the end of every job.
Computer use exists for the apps and sites with no clean API. The docs give the reason in one clause: work *"finishes in the real tools rather than as chat drafts."*
Their first example task, verbatim:
The bot runs up to the login and asks. A research bot named Lincoln does exactly that in Paul Lipsky's 11 August walkthrough, stopping for a LinkedIn account.
Five moments need you, and the docs name every one:
- a password or passkey: the sign-in itself, once per site
- two-factor authentication: the code that lands on your phone
- a CAPTCHA: the check that exists to stop a machine
- a payment or identity check: money leaving, or proof of who you are
- a site that explicitly requires a human: the rule belongs to the site
Everything outside those five runs without you.
When one lands, open *Agent Computer* from the conversation, take control, do only the blocked step, hand it back, and tell the bot to continue from the current page.
Browser sessions persist, so the next task on that site usually starts signed in.
And because there is one browser for the whole account, the session you opened for Vince is already there for Mara, Cole, Rina and Owen.
Passwords go into the secure secret request, where the value is masked and never enters the conversation.
*Sign in once and the whole roster is signed in.*
Use a connector where one exists
Prefer a connector wherever one exists. It is more reliable than clicking through a website, and the browser is there for the sites that never got one.
Adding one takes four moves:
- open *Settings* → *Plugins*
- choose the connector
- press *Add*
- finish the authorization in the browser
Installed connectors are account-wide, so one add covers every bot.
In the message itself, *@* attaches a connector to the task and */* calls a saved skill.
On Hacker News, jjcm had a bot work through roughly 40 fabric factories in Vietnam a month into using the product. They answer on WhatsApp:
Emails are monitored far less. It's one of the reasons I haven't been successful with this in the past despite trying
It reached the first 10 on its own, and jjcm pushed it to the other 30 himself.
That job existed only inside a live account. A chat draft would have been worth nothing to him.
Step 5 · Count what the roster brings back
Five of them worked all week and your bank statement shows a subscription. You never counted a single thing they finished, so the seat still reads as a cost.
Count them the way you count a person, by what each one finished.
Give the roster a plain week of finding buyers, writing to them and answering them. Friday looks like this:
- Mara: 214 companies checked against your buyer list, 38 that fit
- Cole: 38 letters written and sent, 9 replies by Friday
- Rina: 26 images for the ads and the site
- Vince: 61 inbox threads answered, 4 of them passed to you
- Owen: the count itself, posted every Friday at 5
Mara's 38 are Cole's 38, and one of those 9 replies signed the contract.
What the week cost
Then there is what the week cost. Three plans give you access:
- SuperGrok Heavy: $300 a month, bought from xAI
- Cursor Ultra: $200 a month, and the same login opens Grok Bot
- Cursor Teams Premium: $120 a month per seat, assigned by an admin
Each includes weekly usage, and work past that pool bills from model and token cost.
On Ultra the $200 is $46 of seat for the week. On-demand ran to $142.80, so the week cost $188.80 against a signed contract of $1,840.
Set that against the 339 finished pieces above and each one cost 56 cents.
Take the seat price for the week, add what the on-demand line says, and divide by the jobs the roster finished.
The pool and the on-demand line sit in *Settings → Usage & Billing*. Overage arrives on a separate Cursor invoice, and its switch is at *cursor.com/dashboard?tab=spending*.
The meter moves on how much the roster has to look at. @malm_magnus ran an end-to-end playtest of his own game on 13 August:
It works, but screenshot-every-move is burning the weekly Bot pool. Next is to reduce the nr of screenshots.
Every screenshot is a picture the model reads before it can act. Route the repetitive work through a connector and the pool stops paying to look at screens.
*The weekly pool pays by the screenshot.*
Owen prices the week himself, and the whole thing fits in one Friday routine:
Your numbers go in and Friday comes out priced by role.
All five of them live on one computer, and that computer wipes most of itself every time it updates.
The one move that keeps it all alive
Every tool you install goes where tools go, a global install into home. The computer then rebuilds itself on the latest image, and Owen's CLI is not on it any more.
You have more room on that box than anything in the interface lets on. One command, run on a live Grok Bot computer on 14 August:
User *box* carries *(ALL) NOPASSWD: ALL*, so the word between you and root is *sudo*. The docs say the bot starts as a non-root user, and it elevates without a password.
The docs put it softly. Files, browser state and supported sign-ins are *"designed to survive normal computer updates and recovery,"* and you should *"treat temporary directories, manually installed packages, and uncommitted application state as replaceable."*
The same 14 August check is blunter about it. After an update only */workspace* was standing: */home/box/deps* came back empty, apt packages and */usr* were gone, and Tailscale and ssh had to be installed again more than once.
One folder holds the company. Four subfolders cover everything the roster owns:
- */workspace/bin*: the CLIs your bots call, with that folder on *PATH*
- */workspace/config*: every config file those tools read
- */workspace/skills*: the methods you taught, out of any home directory
- */workspace/state*: files, results and project state the work produces
Home becomes a cache from there on. Nothing the roster needs rides on an apt package.
The same line goes at the top of every skill, and it is what puts the tools back:
The roster puts its own tools back before it starts working. A missing CLI reinstalls itself and the job continues.
*The roster is only as permanent as the folder it lives in.*
Prove it on your own computer
Run this on your own computer and wait five minutes:
One of those two files comes back. Everything the roster needs goes in the folder that kept its file.
Nobody has published this. Grok search in comprehensive mode, an X search over 11 to 14 August and a 14-day sweep of my own parser all turn up nothing on surviving an update, and no path map anywhere.
Tailscale, cron jobs and whole CLI setups are going onto these machines this week. @mvanhorn told his bot on 13 August where his skill lands:
That should land at *~/.agents/skills/last30days*
The same install into */workspace/skills* survives the rebuild. *~/.agents/skills/last30days* does not.
You are the manager now
Everybody starts in a chat window, typing a question and reading an answer. Somewhere around the second hire that stopped being the job.
You built three things and wrote all of them yourself: the names, the descriptions, and */workspace*. None of it came with the product, and none of it goes away when the product changes.
You have a payroll, and it is five names long.
*You stopped asking for work and started assigning it.*
Open the app and hire one. A name, a job, and one line about what it owns, and the company exists before the day starts.
If you want to stay up to date with everything happening in AI, follow me on X and Telegram:
X - https://x.com/Argona0x
Telegram - https://t.me/+r0clI4-MMC03ZjAy
