Team Bots, and the Lead-and-Workers Pattern
Once a Bot is useful to you, the next question is whether it can be useful to other people too, and whether several Bots can share a bigger job without turning into a mess. This tutorial covers both. The first half is about Team Bots: how a Bot shared with colleagues in a workspace behaves, and how to set it up so it helps everyone without leaking, drifting or running up a bill. The second half is about a pattern worth copying on a small scale: one lead Bot that plans and several worker Bots that do. You will learn when each is the right tool, how to set the ground rules, how to read the reports, and where the pattern fails.
Difficulty: ★★★☆☆ (some practice with single Bots and job cards helps)
Tools needed: Grok Bot, and for Team Bots a Team or Enterprise plan
Last updated: November 2026
Overview
A Bot that serves only you has a simple life. It has one boss, one set of preferences, and one person who reviews its work. Share it with a team and all three get more complicated. Whose preferences does it follow? Who sees what it knows? Who checks its work, and who pays when it works hard?
The Grok Bot 1.0 announcement describes the feature. You can share a Bot with specific people or groups and use it in Slack or Microsoft Teams. It is for Team and Enterprise plans. A shared Bot learns from everyone it works with. The announcement says it still respects each person's permissions and checks whether something is fine to say in a public channel. It also mentions settings and analytics, including usage, so you can see how a Bot is being used and manage what it knows.
That is a lot of capability, and the sensible way to use it is the same as with a single Bot: one job, written rules, small permissions, and a regular review. The difference is that mistakes now have an audience.
The second idea in this tutorial is about scale. Lauren Tan, an engineer on Grok Bot, describes running more than twenty Bots at once. Her numbers are her own and this is her account, so take it as a story and not a promise. But the pattern underneath is simple enough to copy on a small scale. A lead Bot plans. It breaks a big project into tasks and does not do the work itself. Worker Bots do. Each has one job and its own notes. Each reports back in three lines. And you decide the hard parts: money, public posts and anything hard to undo come back to you.
The tutorial has seven steps. You decide whether a Team Bot is the right tool. You set it up with one job and ground rules. You decide who sees what. You watch usage. You build a lead-and-workers set for a bigger project. You read the reports. And you learn the ways this pattern fails, so you can stop it before it does.
Who This Is Useful For
What You Will Learn
What You Need
1. Grok Bot, and for sharing in a workspace, a plan that supports Team Bots. The announcement says it is for Team and Enterprise plans.
2. At least one single Bot with a written job card.
3. A project with distinct parts, if you want to try lead-and-workers.
4. An hour to plan.
The official notes on sharing Bots with teams, Slack and Microsoft Teams, settings and analytics.
An interview with people who run many Bots, including Lauren Tan. The full list of Bots is behind a paywall.
Information as of 14 November 2026. Plan availability, limits and screen names may change, so this tutorial relies on none of them. Lauren Tan's account of running more than twenty Bots is her own, and is reported here as a story to learn from, not as a measured result or a promise.
Step 1 — Is a Team Bot the right tool?
A Team Bot suits a job that several people ask for again and again, where the answers draw on the same material. Good fits:
Poor fits:
Use the same question as for any Bot: what does it read, what does it produce, what must it never do, and who checks it? If you cannot answer in a sentence each, the job is not ready to be shared.
Step 2 — Set it up with one job and ground rules
The advice from the earlier lesson applies with more force. One job per Team Bot. "Support triage" is better than "helper". A Bot called "helper" will be asked everything and will be wrong in every direction.
Then write the ground rules in the channel where people will use it. A pinned message works well. Say what to ask it, and what not to.
Support triage Bot: ground rules
What to ask it: sort new customer messages, draft a reply, flag urgent ones, find an earlier similar case.
What not to ask it: refunds, legal questions, anything about a named employee, anything with card or ID numbers.
It never: sends anything to a customer. A person sends every reply.
Who owns it: Sam. Corrections and new rules go to Sam, not into chat.
Who reviews: Sam checks ten replies a week.
If it is unsure: it says so, and tags a person.Three details matter.
1. One owner. Someone who answers for the Bot, takes the corrections and decides what it learns. Without an owner, the Bot learns from whoever shouts loudest.
2. A rule about sending. For a shared Bot the safest default is "a person sends everything". Relax it only for a narrow, low-risk case, after a long period without mistakes.
3. A way to say "I do not know". A Bot that always sounds sure is dangerous in a team, because nobody can tell its guesses from its facts.
Step 3 — Decide who sees what
Before you share, decide the access. This is the part teams rush and regret.
Remember how a shared Bot's memory works. It learns from everyone it works with, so anything said to it may shape what it says to others. Tell your team plainly: do not give the Bot anything you would not want another colleague to see surface. Personal circumstances, HR matters and confidential client details stay out.
The announcement says the Bot respects each person's permissions. Test that, rather than assume it. Ask a colleague with narrower access to request something that needs wider access, and see what happens. If it answers anyway, tighten the setup before you go further.
Step 4 — Watch usage and review its notes
A shared Bot can be used a lot, by many people, without anyone noticing the total. The announcement mentions settings and analytics, including usage numbers. Look at them.
Weekly, for the first month:
1. Check usage. Is it used by the people you expected? Is anyone using it far more than the rest? Is anything running in the background that nobody asked for?
2. Read a sample of its work. Ten replies, five summaries, whatever the job produces.
3. Read its notes. Ask it what it remembers, using the method from the memory lesson, and mark each note right, wrong or unsure. A shared Bot is where a wrong note spreads fastest.
Monthly:
1. Review the ground rules. Do they still match how people use it?
2. Review access. Has anyone left or changed role?
3. Review the cost to you. Whatever your plan counts, make sure the total still matches the value.
Do not quote numbers about cost or limits to your team unless you have them from the official pages. Plans and prices change.
Step 5 — The lead-and-workers pattern
Now the second idea. When a project has several distinct parts, one Bot trying to hold the whole thing gets muddled. The pattern from Lauren Tan's account, which is simple enough to copy on a small scale, divides the work like this.
The reason this works is the same reason a small team works. The lead holds the plan and the overview. Each worker holds only what its task needs, so none of them gets confused by the whole project.
A small version for a solo founder planning a product launch might look like this:
| Bot | Role | Job | Never |
|---|---|---|---|
| Chief of Staff | Lead | Break the launch into tasks, track progress, list open decisions for me | Do the tasks, spend, publish |
| Research | Worker | Competitor and audience research, with sources | Contact anyone |
| Outreach drafts | Worker | Draft emails to journalists and partners, in my voice | Send anything |
Start with one lead and two workers. Add more only when you can review what they send back. This is the same rule as for a team of single Bots, and it matters even more here, because the lead multiplies the work that reaches you.
Give the lead a clear goal and clear limits:
Goal: Plan the launch of my new course for 1 December. Break it into tasks and assign each to Research or Outreach drafts. Track what is done. Every evening, give me a short list: what is done, what is blocked, and what needs my decision.
Limits: You do not do the tasks yourself. You do not send, publish or spend anything. Money, public posts and anything hard to undo come to me.
Each worker reports in three lines: what I did, what I found, what I am unsure about.Step 6 — Read the reports properly
With a lead and workers, you read more reports, not fewer, unless you read them in the right order.
1. Start with the lead's evening list. What is done, blocked, and waiting for you. Make your decisions from this.
2. Read the three-line reports. The "unsure" lines first. They show where the real risk is.
3. Open the full output only where needed. For anything you will rely on, open the source. For research, open two sources yourself.
4. Check the handoffs. Did the drafter use what the researcher found, and only that? Did anything get changed on the way?
Keep a short log of decisions you made. When the project ends, it tells you where the Bots helped and where they wasted your time.
Step 7 — Where it fails, and the stop rules
This pattern fails in predictable ways. Know them in advance.
The lead delegates badly. Tasks are too vague, or too big, and the workers guess. Signal: worker reports with long "unsure" sections. Fix: ask the lead to rewrite the tasks with a clear "done looks like".
Workers drift. A worker does something nobody asked for, because the thread suggested it. Signal: output outside the job card. Fix: restate the card, and tighten its permissions.
Handoffs lose facts. A claim passes through three Bots and picks up detail it never had. Signal: a number or quote in the final output that you cannot trace. Fix: require sources at each step, and verify any figure that matters.
Circular work. Bots keep handing the same item back and forth. Signal: the same task reappearing, and usage rising. Fix: stop it.
Cost creep. More Bots, more runs, more usage. Signal: usage numbers climbing faster than value. Fix: watch the analytics weekly.
Write stop rules before you start, and put them in the lead's brief.
Stop and tell me if: a task has been handed back and forth more than twice; any worker reports "unsure" about something involving money, a public post or a person's data; a source cannot be found for a claim that will go in front of other people; or the work is not finished by Thursday.A stop rule is cheap, and it turns a runaway process into a short pause.
Worked Example: a small agency's support Bot
A four-person design agency gets a stream of client questions by email and in Slack. They set up a Team Bot called "Client desk".
One job. Sort incoming client messages, draft replies from the agency's own past answers, flag anything urgent, and find the earlier similar case. Never sends, and never discusses fees.
Ground rules are pinned in the channel, with an owner, Dev, who takes all corrections. Only Dev and one other person may teach it. Everyone may ask it.
Week 1. It is shared with the four of them. They test its permission handling: a colleague without access to the finance folder asks about an invoice, and the Bot declines. Good. Dev checks ten drafts. Two use the wrong tone for a long-standing client, and Dev adds a rule.
Week 3. The usage numbers show most requests come in on Monday mornings. Dev schedules a lighter review for the rest of the week. When Dev asks the Bot what it remembers, one note is wrong: "Clients get a free revision round." That was a one-time promise to one client. Dev removes it.
Month 2. The agency starts a larger project, a client rebrand, with a lead and two workers: Chief of Staff, Research and Drafts. The lead's stop rule triggers once, when the same task is handed back three times. Dev reads it, finds an unclear brief, rewrites it and restarts. The rebrand finishes with the team reviewing about twenty minutes a day.
Common Mistakes to Avoid
Three habits cause most of the trouble with shared Bots and bigger teams.
Mistake #1: Sharing a Bot before it has rules. A Bot with no owner and no ground rules learns from whoever talks to it most. Fix: run the job yourself first, write the card and the ground rules, name one owner, limit who can teach it, and make "a person sends everything" the default.
Mistake #2: Trusting that permissions and public-channel checks will protect you. The announcement says the Bot respects permissions and checks what is fine to say in public, and that is a safety net, not a plan. Fix: test with a colleague who has narrower access, keep sensitive material out of the Bot, and tell the team not to say anything they would not want surfacing for others.
Mistake #3: Building a big team of Bots without stop rules. A lead that delegates badly, workers that drift and handoffs that lose facts all grow quickly. Fix: start with one lead and two workers, require three-line reports and sources, keep hard parts with you, and write stop rules before the first run.
Going Further
Rotate the owner. Every quarter, a different person owns the Team Bot for a month. Reading the notes with fresh eyes catches habits the usual owner no longer notices.
Split a Team Bot that has grown. If one shared Bot covers support, onboarding and sales questions, divide it into three. Each gets a card, an owner and its own notes.
Compare to a human process. Ask what the work looked like before: who did it, how long, how often it went wrong. Use that as your yardstick for whether the Bots help.
What to learn next. The next lesson goes through real jobs in detail: inbox, sales outreach, expenses, research, meetings and documents. It gives you a brief and a set of checks for each one.
Key Takeaways
In this tutorial you learned:
Before you move on, a challenge. Choose a job that two or more people ask for repeatedly, write the ground rules and the owner, and decide who may teach it. Or, if you work alone, choose a project with three parts and write the lead's brief with stop rules. Run it for a week and write down one thing you would change in the rules.