Grok Bot 201: a SpaceXAI engineer's multi-bot setup, explained

<aside> 🎥

Source: Grok Bot 201 on the official Grok YouTube channel. Published Oct 21, 2026. It's a recording of a live session from Oct 9, about 55 minutes long. Speakers: Joseph Yang, Senior Field Engineer at SpaceXAI, who runs the demo, and Scott Lerner, Senior AI Deployment Manager at SpaceXAI, who hosts. Everything below comes from the video's auto-generated captions. Quotes are verbatim, with filler words like "uh" cut and marked with "…". Model names and numbers are written as they appear in the captions, and anything uncertain is marked. Timestamps are approximate.

</aside>

TL;DR

Joseph runs a small team of specialized Grok Bots. Roy is the orchestrator, Ed handles metrics, Dex designs and Lou writes code. Together they maintain a Grok Bot sticker shop website.
He gives each job its own bot, the way a startup gives each role its own person. Each bot has "their own computer, they have their own context. Context is king in this era, and they have their own memory as well."
A routine runs every 24 hours. Roy gets page-view metrics from Ed, drops the bottom 3 stickers, has Dex design 3 new ones, and has Lou update the site. He calls the result a "self-evolving website."
For design work he uses a fan-out: one prompt sent to Cursor cloud agents running 3 different models. He compares the screenshots and screen recordings and picks what he likes, taking ideas from each.
Token tip: get your prompt right on a model you know well first, and only then fan out. Vague prompts "could be interpreted in different ways" and waste tokens.

<aside> ⚠️

Scale check. This is a demo project. Joseph built the sticker shop in "maybe… two hours in total." The metrics come from Vercel analytics. Before the session went public, only "eight out of 24 stickers" had been clicked, and the top one had 5 page views across about 2.5 hours of data. He also says the setup took real iteration. He spent "almost 30 minutes" with Ed alone to get the flow working, and about as long with Lou and Dex. "It's not something that's super easy that just goes ahead, snap your finger, everything works."

</aside>

The setup: 4 bots

Bot	Role	Connections (as stated in the video)	What it does in the demo
Roy ⭐	Orchestrator and primary bot (starred in the sidebar)	The slide shows Roy linked to X and to the website, and it coordinates the other bots	"This primary bot basically has all the context, all the information about all the other bots." It runs the PMF workflow, which Joseph built as a skill. Its goal is to "help me find PMF of these stickers."
Ed	Metrics bot	X and the website. Ed signs in to Vercel in Chrome on its own computer. He mentions a Vercel API token as another option.	Pulls page views and unique visitors for each sticker from Vercel and shows them on a live "popularity canvas," a dashboard feature he calls not yet released
Dex	Sticker designer	Figma. Dex is a team bot built and shared by Martin, a field engineer.	Makes new sticker designs, for example "a happy blue one" made live during the demo. Grok API is mentioned for producing the assets.
Lou	Full-stack developer	GitHub, "so that it can push code for me"	Builds the site and swaps stickers in the code base. The diagram shows a deploy to Vercel. Lou has its own computer with a terminal and can run localhost.

Two more bots show up in the video but are not part of the daily loop:

Sam, a UI/UX designer, helped build the original site. In the workflow diagram Joseph says Sam "actually doesn't have any part in the PMF loop."
The X bot posts to Joseph's X account. At the start Roy tells it to post "we are live with Grok Bot 201," and you can watch it control its own computer to do so. (It's unclear whether this is the same bot as Ed.)

Why several specialized bots instead of one

Asked "Why not just have one bot that just controls it all?" (around 18:30), Joseph compares it to staffing a startup (around 18:47–19:56):

"If we think about modeling a startup team… if you just have all all the tasks that you want and… you delegate that to just one person to do it, I don't think that outcome will become great… if that outcome is always good, then we wouldn't have job titles today."

He agrees that with AI the lines between roles are blurring, but still keeps them apart:

"You still want to delegate them separately because they still have their own computer, they have their own context. Context is king in this era, and they have their own memory as well."

Scott adds that "maybe in the future we will just have one bot," but for now specializing works best. His advice: "think of it as… like you would teammates. They're specific to their role and responsibility, and they become really good at that."