Before You Build a Grok Bot Team, Check What They Share

0:00Guys, you can build a Grok Bot team and still end up with chats that all pull you back into the work because none of them knows exactly what it owns, what it must return, or when it has to stop and ask you. [music] I tested two fictional bots. A chief delegates a public brief,

0:15then I review its source-linked return. By the end, you can write one role contract, give it a bounded job, and know when to stop it. You will see the delegation, its source-linked return, and the human [music] stop. Start with one job that only uses public or fictional data. The two North Star demo bots stayed inside that controlled scope

0:34and received only three public pages. The official Grok Bot launch page says bots share computer of their own, and I'll show you what this build exposed on screen. Hey, I'm Johnny Nell. I spent years building businesses the old hard way, managing teams, scaling operations, and doing everything manually. Today, everything I build, from automations and no-code infrastructures to fully AI SaaS products, has AI helping in one way or

1:00another. Excuse me, let me quickly show you this. In South Africa, we call this a braai. I build with AI tools that, in my opinion, anyone can learn to use. My mission is simple, to help creators, entrepreneurs, and business owners get clarity with AI. Let's build smarter together and enjoy this video. Before we build the team, I want to clean up one thing

1:22because it changes how you set this up. The official launch page says the bots share a computer of their own. In the Grok Bot viewer that I recorded on August 22nd, 2026, I opened the computer view for the North Star chief, then opened it for the North Star specialist. Both exposed the same non-secret VNC host and path, and these two sanitized desktop crops measured 0.99606 against each other. On screen, you're looking at the same visible computer surface. Now, I need to be very precise about what our screen comparison proves for you. It proves both demo bots expose the same visible surface. SpaceX AI's current docs go further. They say all

2:04your bots share one persistent computer, including files, browser sessions, and app logins. And a bot's screen is not a separate security boundary. We did not open any private resource to test that documented state ourselves. So, you should treat the official shared state warning as the boundary. You are still looking at no evidence about granular permission controls. We did not sign into anything. We did not go looking through an existing workspace. In my opinion, that distinction is where

2:33people get careless with a team setup. And you might be thinking, "Johnny, why does that matter if I'm only asking a researcher to read public pages?" It matters because the title you give a bot does not make its computer separate. The boundary has to be in the job you handed and in the information you refuse to put in front of it. So, for this whole video, North Star stays fictional and the browser stays empty. The only material these bots are allowed to touch is three public URLs. That gives us a clean test without making up a privacy claim the product has not proved. So, let me quickly show you the part that

3:08makes this a team instead of two fresh chats. The first bot became North Star demo chief and received a narrow assignment. I told it to take one public research brief delegated only to the North Star public sources specialist and bring the artifact and the source links back to me. I also told it not to touch existing workspace material or a business workflow. The bot automatically renamed itself after it accepted that contract. And you can see the receipt on screen now. I then created the specialist for public source research in this demo only. Its contract on screen requires a labeled return that makes uncertainty visible. If a source was

3:46missing, it had to say that. If the three pages disagreed, it had to surface the disagreement. I told it not to widen the search or act beyond the approved reading. Dude, writing these role cards feels slower than typing one big prompt, but a role contract gives you something much more useful than a clever description. It gives you a way to read the result and say whether the bot

4:07stayed inside the work you handed it. For your own first team, keep the contract plain. Describe the input and the return, then state the judgment that remains yours. For tests like this, one job that ends with artifact gives you enough to review. Now, the chief has a job, so let's give it one. The work packet you see is called the North Star Studio Weekly Public Agent Brief. It contains exactly three pages: the XAI Grok Bot announcement, the Corsa pricing page, and NIST's AI Risk Management Framework page. The request says, "Do this research yourself, then return concise notes with source links in the labeled review format." After reading

4:45those pages, the specialist must stop. Instead of disappearing into a long answer, the chief sends the packet to the named specialist, and the exchange becomes visible between the two bots. The specialist receives the exact packet, which you can read without any extra instruction added. The bot exchange is read-only, so I can open the conversation and read the handoff rather than hoping the chief paraphrased it correctly. You can inspect the packet for yourself rather than take a handoff on trust. And you can see the same boundaries the specialist received. The useful move is opening the exact packet the specialist received, because that tells you whether the bot is solving the

5:21job you gave it or version it quietly invented for itself. And you might be thinking, "Johnny, can't I simply ask one bot to do all of this?" You can, most probably for a small task. I split this one because the point of the demo is the handoff. The chief sends the written boundary to the specialist, the specialist does the reading, and then I see the evidence come back through the same path. And guys, if you You this set up ready to copy into your own first control test. The GrokBot team contract

5:50pack is inside the AI founder's vault. It has the fictional workspace brief, the chief card, the specialist card, the handoff card, the review sheet, and the list of product questions this demo did not answer. You also get the community with the other builders working through this stuff properly. The link is down in the description below. So, let's look at what the specialist returned because that artifact is what I can review. The specialist came back with the source linked page you see in the exact format we asked for. It had a confirmed section, an unclear section, and a next check section. The chief then brought

6:27that page back to me in its own chat. You can see the full return on screen, which is why I asked for a visible artifact instead of a friendly summary in a chat bubble. The confirmed material was straightforward. The specialist could point to the three pages it had read and say what each one covered. The useful part came in unclear. On the product pages in this run, GrokBot

6:49eligibility did not reconcile cleanly. The bot did not get to settle that by guessing. It put a human verification of the live cursor surface in next check. So, you can see the missing check instead of a confident guess. That is a much better result than a bot acting certain because it found words that sound close enough. It kept the gap inside the artifact where I can see it instead of burying it in a sentence that looks finished. Now, I'm not going to repeat product details the bot generated from this screen. You verify those against primary sources before you repeat work based on them. The bot produced artifact proves the handoff,

7:27not an audience-facing product claim. You can compare the source links with your original packet before you repeat the job. The proof here is what we watched. Role contract delegation, bot exchange, returned evidence. The unresolved source conflict tells you the job is not ready for another run. So, I read the result against the original packets and the answer was no. Do not repeat this job yet. The artifact did what we asked by showing the three source links and naming what it could not settle. But, the two product pages still did not reconcile the eligibility question and that is enough to stop a public source routine before it becomes

8:03a recurring little machine in the background. I told the chief not to create a routine, not to search beyond the three approved pages and not to inspect anything already living in the workspace. The chief came back with one terminal result, stopped, human check required. Guys, that line is the result I wanted. You can get a tidy document that still carries an open question. If you schedule it anyway, your next run repeats the same gap and gives you a cleaner looking answer for the same unresolved problem. The human review is where you decide whether the evidence is good enough to repeat. Change the contract or stop the whole job. And you

8:41might be thinking, Johnny, why stop when the bot followed the format so well? I stopped it because format is not verification. The specialist returned a sensible next check and the chief respected the stop. The first control test gave me a sensible next check and a stopped job you can verify on screen. You can see two controls in this build that we deliberately did not run. In the computer view, you can see a teach a task control. In the details area, there is create routine. Both remain visible but unrun because the human review had already rejected the artifact. Teaching the disputed research path or scheduling it would have turned a visible conflict

9:21into a repeatable mistake. This is also where I want to draw the line around the shared computer proof one more time. Our capture proved the chief and specialist exposed the same visible VNC surface. The official docs say all bots share files, browser sessions, and app logins on that computer. We do not place anything private there to test that state ourselves. And we did not test granular permission controls. You should keep the screenshot in the documented claim separate rather than use one as proof of the other. So, when you open Grok bot and see a routine button, do not make the first question, "What can I schedule?" Ask whether the finished

10:00artifact has earned a second run. If it has not, you have your answer already. You should leave the control alone and fix the work back at first. One more thing, because a chief and a specialist can make you want to build a whole fake company before lunch. Keep your first team small enough that you can inspect the entire path. Give the chief one

10:19named specialist and one source back. Ask for one finished artifact, then make one human decision. You are learning what the current product shows you, not trying to hand over a business on the first afternoon. For North Star, the job was deliberately boring. Three public URLs and one concise page back. There was no client list, no inbox, no sign-in, no purchase, no send, and no routine. That is why we could see the weak points without creating a mess we then had to unwind. The official

10:49announcement calls Grok bot early beta. Treat the parts you cannot inspect as open questions, then let the work package show you what your first test still needs. Big takeaway for me was this. The first useful Grok bot team is not the one with the most names in the side bar. It is the one where you can point to the role contract, read the delegation, open the returned evidence, and make a human decision before routine exists. For our challenge, this is exactly the kind of build I want to keep doing. We put a real product through a small controlled piece of work. We keep the evidence visible, and we stop where

11:22the product has not earned more trust from us yet. I can't say too much as of yet, but the next part of this journey is going to be taking a verified work package and seeing what changes when the source question is genuinely closed. Which recurring job could you test with public or fictional data before it touches a real account? Drop it down in the comments because I'll be reading through them. And guys, if you want the North Star brief, the chief and specialist role cards, the handoff card, and the review sheet, they are inside the Grok Bot Team Contract Pack in the

11:52AI Founders Vault with the community. The link is down in the description below. If you feel called to it, subscribe to follow the journey. See you in the next one.