Logos52
wiki / Systems / AI and Agentic Systems / Grok Bot Primer

Grok Bot Primer

concept updated 2026-08-31

Grok Bot Primer

Grok Bot comes with a paid subscription to its maker’s AI service and gives you a computer in the cloud that keeps running after you close your laptop. On that computer you set up named helpers, and each helper is a small program with a job description, a conversation you talk to it in, and a schedule of tasks that the product calls routines. The people selling it and the people showing it off promise a staff that works while you sleep, where one helper reads your inbox, one renews your subscriptions when a card expires, and one runs your business.

One fact decides how a person should use it, and that fact is easy to miss inside that promise. Your account gets one cloud computer, and every helper you make lives on it, so the helpers share its files, its browser, and whatever you have logged into on it. The subscription also comes with a weekly allowance of usage that every helper on the computer draws from, so what one helper spends is gone for the rest. Each helper gets its own screen, but a screen is only a place to work, and a helper on one screen can open a site that another helper logged into on a different screen, which means a password typed for one job is there for every job.

Because of that, the question about each helper is what it should be allowed near, rather than what it can do. Take the owner of two public websites, a Chinese dictionary and a graded reader, which is a site of short Chinese stories sorted by level with the same story characters running through them, plus a public notes site, and a habit of reading about how other people work with AI. If that owner turned Grok Bot on tomorrow, what would he set up, what would he refuse, and how would he know the setup was working?

One helper, one job

Every helper in the owner’s setup has one job, and the job is a standing duty rather than a single task. One helper, named Watch, looks after the public websites. It has a clock, which runs it three mornings a week, on Monday, Wednesday, and Friday, a file it writes when it runs, and its own chat, where it posts what it wrote in the file, and whatever a helper posts in its chat shows up on the owner’s phone as a notification from that helper. Its description ends with a line saying what it never does, which for Watch is publish a new version of a site, change a site’s settings, or ask a search engine to index anything. One job per helper is the rule because a helper with one job has one output, and one output is easy to look at, so in the week the helper quietly breaks, the file is late or empty and the owner sees it, whereas a helper with four jobs can break in one of them and still look fine.

Each helper’s output is a report, and a report is all that any helper produces. No helper in this setup changes anything the public sees. None of them publishes a page, accepts a change to a site’s code, or writes to the notes site, because a helper that fixed what it found would be checking its own work, and because the notes site is public and carries the name of the person who owns it, so any edit a helper made would be an unreviewed publication under that name. A report can be wrong and cost nothing, and an edit cannot be wrong without costing something.

Watch has every part of that shape and follows that rule. It checks six public addresses, which are the front page of each website, a blog page on one and a browse page on the other, the front page of the notes site, and the build job that publishes the notes site. It expects each to answer, it also reads the week’s visitor figures for the two websites and a search engine’s report on which of their pages the search engine has indexed, and it writes one file with all of that. If everything answered, the file says so, the message in Watch’s chat is one line, and nothing else happens, so a quiet Monday morning means the sites are up.

The helpers and the empty middle

Watch is one of the helpers that read the world and report on it. A helper named Field reads a public feed of write-ups in which named people describe the helpers they run, and it looks for the one item in a week that would change a file, a setting, or a routine in this setup. A helper named Recap reads the captions of new episodes of the podcasts the owner follows, the written text of what was said that a show publishes with its video, and writes a summary with timestamps and quotes, because the owner wants what was said without listening to the whole show. Three smaller reading helpers do narrower jobs on the same pattern. One keeps a queue of Chinese articles worth reading for practice, one keeps a calendar of update dates for three online games the owner plays, read from the games’ public pages, and one audits the notes site once a week for dead links, duplicate pages, and pages without sources. Every reading helper has the same shape as Watch, which is one job, a clock, a file, and a line saying what it never does.

Brief is a different kind of helper, because it does not read the world. It reads what the other helpers wrote, plus a short list of the changelogs for the tools the owner works in every day, which are the AI assistant he writes and codes with on his own laptop, a notes app, and Grok Bot itself, and every weekday morning it sends one notification to the owner’s phone of at most five lines, with the most important line first so that the notification makes sense on its own. On a morning when nothing changed, the notification says so in one line.

Steward is the helper that looks after the other helpers, which is one job with two parts. It reports which routines exist, when each last ran, what each wrote, and how much of the week’s allowance each has spent, and on weekdays it backs up every helper’s files. Steward changes nothing either.

One more part of the setup is not a helper at all. When Watch reports a fault on one of the websites, the owner reads the report and decides whether it is worth fixing. If it is, a separate coding agent, a different program from the assistant on the laptop and one that runs on a machine of its own that its service provides rather than on the shared computer, opens a pull request against the site’s code, which is public. A pull request is a proposed change the owner can read in full and accept with one click, and accepting it is called merging. So a helper only ever finds a fault, while the fix is written elsewhere, and the owner approves the fix before it starts and merges it when it is done.

The helper missing from that setup is the one most people put first, a chief of staff in the middle that holds the logins, takes requests, and hands work to the others. The how-to pages published by the maker of Grok Bot put that helper first: a sales playbook, a product-manager playbook, and a six-seat studio write-up all start with a manager that routes work and holds mail, an ads account, or a store listing. Those pages are field evidence of what the product is sold to do, not a roster this account copies, because the logins those managers need are the logins this setup keeps off the shared computer. This setup has no such helper, because a middle helper has to hold every login to route every job, and on a shared computer every other helper then holds those logins too. A middle helper also spends its day talking to other helpers, and that talking is charged to the week’s allowance, so that when another user of the product ran a chatty chief of staff, the chief of staff used up a week in a few hours, and a quieter version the same user then built of the same helper used about fifteen percent of that. So the middle in this setup is the owner with a phone, opening some notifications and ignoring others.

How the helpers talk to each other

The helpers never talk to each other in conversation, and everything they pass to one another goes through files. Watch writes the results of its check to a file in a shared folder, Brief reads that file the next morning, and neither helper has a conversation with the other. Files are the rule because a helper’s routines run inside its conversation, and a long conversation makes every run inside that conversation cost more, which is why the product’s own team says to keep a scheduled job on a fresh helper with a short history. A file also lets the owner open it and see exactly what was passed, which a conversation does not.

The files also carry the helpers’ work off the cloud computer. Steward backs up the shared folder to a public code repository on weekdays, leaving out the one folder that holds the helpers’ keys, the access codes a helper uses to read a service such as the websites’ visitor figures. That repository holds only the helpers’ own reports, so the backup is the one thing a helper writes outside the cloud computer, and nothing on the websites or the notes site comes from it. A copy of that repository sits on the owner’s own laptop, which is a Mac, so one command on the Mac, a pull, copies the week’s reports home without anyone pasting anything. Once a week the owner opens a session with the assistant on the Mac, the one he writes and codes with, reads what is new, and talks through what mattered with the assistant. Anything worth keeping goes one of two ways. If it is knowledge, it becomes a page on the notes site, and if it is a way of doing something, it becomes a skill file for the helper it belongs to, which is a short written method the helper follows when it runs, so that a method Field found this week is something Watch or Recap can run next week.

A week

On Monday at eight in the morning Watch checks its six addresses and writes its file. On Tuesday at seven Brief reads the file, checks the changelogs, and sends its notification, and on a normal Tuesday that notification is one line saying nothing changed. On a Thursday the assistant the owner writes with ships a new feature, and Brief’s first line that morning says so, with a link to the release note. On a Saturday a person the owner follows publishes a guide with a method worth copying, and Field posts one line with the link in its chat, which is its only post that week. On a Wednesday night Recap finds a new episode of one of the shows and writes its summary. Every weekday evening Steward backs up every file. On the other mornings Brief’s one line says nothing changed, and Watch’s Wednesday and Friday runs find the sites up. On Sunday the owner pulls the copy to the Mac, reads the week with the assistant, and keeps the guide and one paragraph of the podcast summary, so that the guide becomes a skill for Field and the paragraph becomes a line on a page. Every other hour of that week the helpers did nothing, and that is how it should be, because a helper that runs when nothing changed is spending the week’s allowance on nothing.

What stays on the Mac

Some work never goes near the cloud computer, and the line is drawn by the same fact as before, which is that whatever is on that computer is there for every helper. Anything that needs a login or touches money stays on the Mac. Mail, cards, shopping, and the owner’s finance dashboard are never set up on the cloud computer, however many write-ups show a helper renewing subscriptions overnight. All authoring stays on the Mac too, meaning the prose on the notes site, the example sentences in the dictionary, the story characters in the graded reader, and any rewriting of Chinese, because publishing is the owner’s act and the helpers only report. Speech is generated on the Mac as well, for a small audio app the owner built for himself to listen to while he exercises, and no helper runs text-to-speech. Decisions about what is good enough to publish, and the settled rules the setup runs under, stay with the owner for the same reason as the authoring.

Two smaller things are handled from the Mac as well. When one of the websites, whose code is public, needs a file that says where the site is going, so that a helper, once one is given that job, can sort the requests and bug reports strangers file against the site into ones that fit and ones that do not, the owner writes that file on the Mac, puts it in the site’s public code, and the helper only reads it. And the keys Watch needs, which give read-only access to the visitor figures for the two websites and to that search engine’s report on which of their pages the search engine has indexed, are minted on the Mac fresh for that one job and only then placed on the cloud computer, and nothing is ever copied from the keys the Mac itself holds for the same services.

What it costs and when to stop

The cost is a weekly allowance of usage that comes with the subscription, and the setup is built to stay inside that allowance. Overage past the allowance can be bought, and in this setup the switch for buying it is off. A single routine run that finds nothing new costs about one hundredth of one percent of the week, so twelve or fifteen runs a week are nothing. Three other things are what empty a week. The first is a long conversation, which makes every routine on it cost more. The second is a routine that fires every fifteen minutes, which is about a hundred runs a day, and a different user reported that two helpers polling that often used fifteen percent of a week in half a day, since a run that reads and writes costs many times a run that finds nothing. The third is helpers talking to helpers, which is how the chatty chief of staff described above spent a week in hours.

Four rules decide when to stop. A week that closes above about forty percent of the allowance means cutting back before adding anything. A reading helper whose finds go three runs in a row without changing what the owner reads or does is retired, and retiring it means the setup is working rather than failing; Watch is the exception, because a quiet Watch means the sites are up, so Watch is judged by whether its rare faults get fixed. A helper the owner never opens is spending the allowance for nothing, so it is paused rather than hidden, because the product lets you hide a helper from the list without stopping its routines. And a new helper is added only when Steward’s reports show a gap that no existing helper can cover, and never because a helper sounded like a good idea.

After the first two weeks, five checks show whether the setup is working. Field posted at least one line the owner opened. Brief or Watch was opened from the phone on three weekdays. No routine ran more than a few times a day. Usage stayed inside the week. Every helper has the line in its description that says what it never does. If the first two of those checks fail, the routines added in those two weeks are retired, apart from Watch’s, because a helper whose reports nobody opens is doing no job.

What a setup like this can grow into

Taken as a whole, Grok Bot on this account is a set of reading helpers with no helper in the middle. Watch, Field, Recap, and the smaller reading helpers look outward and write files. Brief reads the files and sends one notification. Steward watches the reading helpers and backs up the files. The only place everything meets is the owner’s phone, and the only thing that changes on the sites is a pull request the owner chose to merge.

Five kinds of work fit a setup like this: watching, repair, keeping, a learning loop, and skills made from what the reading helpers find. Watching and repair are already running, since watching is what the reading helpers do and repair is the pull-request lane. Keeping is the work Steward and the weekly audit of the notes site do now and can grow into, which is backing up files, listing what could be deleted without deleting it, and checking the notes site for dead links and pages without sources. A learning loop is the next reading helper worth adding, and it would be a helper that turns a public paper into a five-minute read or into a script the owner’s audio app can speak, together with a rule that Field and Brief compare what they sent against what the owner opened each month and propose changes to the rules they pick lines by. Skills made from what the reading helpers find is the weekly session on the Mac, where a method Field found becomes a file a helper runs. Two more kinds that other people run are kept off the helpers. Making, in the sense of producing content, happens on the Mac, and the only content that reaches a public site by way of any agent is a proposed change the owner reads and merges himself, the way he merges a fix. Advising, in the sense of a helper that questions the owner or plans with him, stays a conversation with the assistant on the Mac, because on the shared computer that job turns into the chief of staff this setup left out.

So, for the person with two public websites, a notes site, and a reading habit, the answer is to set up a few helpers with one job each, refuse anything that needs a login on the shared computer, and judge the whole thing by which notifications you open.

Open Questions

  • After a month, which notifications still get opened, and which helpers does that retire?
  • Does a learning-loop helper earn a seat before the second month, or does the weekly session on the Mac cover it?
  • Does the weekly session on the Mac keep happening once the novelty is gone, or does the backup repository fill up unread?

Sources