Skip to content
View original post on X: Alex
Published
49/100

Grok Bot builds AI engineering teams with specialist bots and cloud agents

AISpaceXSI engineer Lingxi Li runs more than 200 Cursor cloud agents at once through a fleet of specialist Grok Bot engineers, up from 15 managed by hand. The guide's 10-step roadmap starts with one bot owning one outcome, then adds specialists by domain, a PM bot, and handoffs between bots.
AlexVerified on X
@de1lymoon
View on X

Grok Bot just became an engineering team

One SpaceXSI engineer's bots now run 200+ cloud agents at once, and the same app can call businesses, run your Shopify store and work the night shift

Here's how to build it in 10 steps: https://x.com/i/article/2114636391988539392

How to Build an AI Engineering Team With Grok Bot

Most people use Grok Bot like a chatbot with a laptop. The people who build it run it like an engineering org.

Lingxi Li, an engineer at SpaceXSI, used to manage 15 Cursor cloud agents at a time by hand. Lingxi's bot fleet now runs more than 200 at once. Lauren Tan, a principal engineer on Grok Bot, shipped 2,000+ PRs in a single month.

The difference is not the model -> it's the org chart.

This is the 10-step roadmap to build that org: specialist bots, a manager, cloud agents that write the code, a merge line - plus the non-code hands that turn a script into a team: phone calls, Shopify, its own inbox, X and Slack.

No code. Every step ends with a prompt you can paste into your bot today.

Read the whole thing if you run any engineering work. Almost every step below comes straight out of it:

01. Start with one bot that owns an outcome

A Grok Bot is not a chat window. Matt Palmer, who works on Cursor and Grok Bot at SpaceXSI, sums it up as an agent with a computer: each bot runs on a persistent cloud machine, so work keeps going after your laptop closes.

The underrated part: show a bot a workflow once, and it saves the steps for next time.

- The official docs give the hiring rule in one line: give one bot ownership of an end-to-end outcome, and add another only when the work has a stable specialist role.

So don't spin up five bots on day one. Hire a Tech Lead. Its job is to be the bot you message when you don't yet know who should own a task.

The rule: one bot, one outcome. The roster grows only when this bot's memory starts holding two jobs.

02. Staff the roster: one domain per bot

Lingxi runs five engineer bots, and each one owns a single area:

- Baltata - the mobile shared layer and everything iOS

- Shaoruru - the desktop client and CI/CD

- Hogan - infrastructure, plus user issues with unclear ownership

- Craig - Android

- Quill - the Grok Bot harness

They can work across each other's areas. But each has its own memory and limited context, so the specs and design principles stay sharper inside the area a bot owns.

- You don't have to write every bot from scratch. The marketplace has ready-made engineering hires: Lingxi's Engineer Bot, a Nightly Audit Engineer, and a SWE bot from Cursor. A template copies the bot's skills, memories and official plugins into your account.

Watch it before you write a single charter. Copying a working bot and tuning it beats starting from a blank page.

Name each bot after its job, then give it a charter. Same structure for every hire, so they all know where the lines are.

The rule: specialists beat generalists. A bot that owns one area remembers it better than a bot that owns everything.

03. Hire a PM and wire the handoffs

Five bots in five separate chats is five assistants. Five bots that pass work to each other is a team.

Grok Bot has two ways to wire that. A group chat holds two to six bots around one outcome: write normally and the bots decide who answers, or @mention one to give it ownership. Bots can also message each other directly - the receiving bot wakes up, handles the request and replies later.

Eric Zakariasson took it one step further. Each project gets its own channel and one row in a Notion database. A PM bot proposes the roster, reuses existing bots first, and caps each channel at six. When a bot gets stuck, it marks its task Blocked and pings him.

His PM rules are worth stealing as-is: reuse the bench first, five specialists max, and no new hire without his yes.

That pattern is now a marketplace bot called Projects Manager. Or write your own:

One catch from the docs: approval requests, login requests and email or Slack drafts don't show up in a group chat. Each bot sends those to you directly - so watch your DMs, not just the group.

04. Hand the code to Cursor Cloud Agents

Here's the shift most people miss: your engineer bots shouldn't write the code themselves. They should manage the agents that do.

A September 30 update made this the core path: bots hand coding tasks to Cursor, manage PRs through GitHub and Origin plugins, and share video demos of what they built. Each Cloud Agent gets its own VM, checks out the repo, does the work and opens a PR. Your bot stays free to manage.

Lauren's favorite setup: a team engineer bot in Slack that you @ whenever there's coding work. You can also ask it to start a Project - a group of related agents in one conversation - and use any model available in Cursor.

That last line is the whole idea. Each agent has its own computer, so your bot stops coding and starts managing.

- The trick is in the brief. Lingxi's bots launch every agent with a thorough prompt that spells out what must be done and what proof is expected. No proof, no done.

05. Close the feedback loop and draw the merge line

Lingxi calls a complete feedback loop the key to keeping a bot team running. Cloud agents take screenshots, the bot checks them with its own eyes, and pushes back if the result doesn't match the ask.

Chat scrolls away, so the loop lives on a board. Each of Lingxi's engineer bots runs a shared Notion database and sweeps it every 30 minutes. For every PR it checks three things: Bugbot comments or security findings, failing CI, and merge conflicts. Anything off goes back to the agent and the row moves back to Working.

Then comes the merge line. If the review is highly confident and the blast radius is low, the PR merges on its own. Everything else waits for a human.

The rule: decide the merge line before the first PR lands. Lingxi compares it to self-driving - trust is built area by area, not all at once.

06. Give the team a phone

A real team doesn't only live in the repo. It talks to vendors, books calls, chases support tickets. As of this week, your bots can too.

Calls on your behalf. On October 25, Elon Musk posted that Grok Bot can now call a person or a business for you, make the request, take notes on the answer and send you a summary.

Over 2.4M people saw it in a day. The official @bot account still hadn't said who gets it, or when.

His post quotes notes on the 0.77 build: the bot dials, sits on hold, presses keys and gets transferred while you watch a live transcript. SpaceXSI hasn't documented the feature yet, so not everyone may see it right away.

Before native calls, creators were already wiring it up. Ole Lehmann showed how to give your bot its own phone number through Bland - the bot sets up the account itself, you just pick the area code and confirm a code.

Think refunds, restaurant bookings, the gym membership you keep meaning to cancel - anything a website won't let you do.

- Your phone's apps. Since October 22, Grok Bot can use Calendar, Reminders, Contacts and more apps on your phone, rolling out on iOS and Android.

- One button to your bot. Matt Palmer's 7-step setup puts Grok Bot on the iPhone Action Button: press it and you're in a voice chat with your bot.

A minute to set up. After that, your whole team is one button press away.

The rule: a bot on the phone speaks for you. Put what it may not agree to in the prompt, every time.

07. Plug it into the business: Shopify, shopping, its own inbox

The code ships the product. Someone still has to run the store, buy the tools and answer the email. Three October features hand that to the team.

Shopify. On October 8, the official @bot account announced that Grok Bot can help run your Shopify store: check orders, track inventory and keep product listings up to date.

9.8M views, and Shopify's own help page backs it up. One heads-up: inside Grok Bot, answers come back as text only.

Setup is one message. Musk added that you can just ask your bot to add the Shopify connector. Or open Connect apps in Grok Bot and search for Shopify - Shopify's help page notes the Grok connectors page does not install it for Grok Bot.

- Shopping. Since August 28, bots can buy things online through Link. The 0.68.1 build notes also mention approving spending with a passkey. Use it - Musk's own tip of dropping a photo of your card into the chat comes with no published spending limits. Also, some users report Amazon blocking AI agents from placing orders.

Its own inbox. Since October 9, a bot can claim its own address on mail.grokbot.com to sign up for services, contact businesses or schedule meetings. Lauren's tip: just ask your bot, or tag @bot on X.

Can't tag @bot? Sign in to grok.com with your Grok Bot email, link your X account, then try again.

Mail it sends waits for your approval on a draft card. On Cursor teams, an admin has to switch on Agent Email first.

The rule: the money side gets the strictest approval line on the team.

08. Make X and Slack the team's front door

Bugs don't arrive in your bot's chat. They arrive in Slack threads and X replies. So put the team where the work lands.

Slack. Team Bots, launched September 28, are shared bots with their own plugins, skills and credentials. You add one to Slack with one click, and each thread becomes its own conversation. Lauren's take: Grok Bot is now multiplayer.

Give it the plugins, skills and credentials for one role, drop it into Slack, and the whole team can work with it.

X. Since October 7, bots can search, read and monitor X natively. You can also reply to any post with @bot and an instruction - the task goes straight to your own Grok Bot. Lauren used that to set up a 9am Slack digest of the best Grok Bot ideas on X, with suggestions for her own workflow.

Steal the shape of that request: a source, a time, a channel, and what to do with what it finds. One sentence, one routine.

And as of Matt Palmer's October 20 update, a bot can also read and reply to your X Chat DMs. He says limits are in place so it can't spam people.

One cost detail from the docs: when a teammate talks to a Team Bot, the usage counts against that teammate's own allowance, not the owner's.

09. Hire an ops bot and run a night shift

Every team drifts. Bots forget complex workflows, repeat mistakes, and new hires start cold. Lingxi's fix is a head of operations named Jenny - the only bot on the team that doesn't write code.

- Daily 1:1s. At 5 a.m., Jenny meets every bot to review the playbook and surface blockers.

- Postmortems. When a bot makes a mistake, it goes to Jenny for root cause. Jenny updates the playbook and tells the other bots.

- Onboarding. New engineer bots get created, briefed on the team rules and introduced by Jenny.

Then there's the night shift. At 3 a.m., Lingxi's bots clean the codebase: dead logic, load time, bundle size. Every morning there's a fresh set of cleanup PRs. Lauren's launch post pitched the same idea: bots that keep working while you sleep.

That was day one. She was already using it to manage her calendar and ship code.

10. Make it a real team: permissions, sharing, review queue

Up to now, the team works for you. The last step is making it work for your whole company - without handing everyone the keys.

Shipped. Team Bots get published for the whole team; until then only you can see them. The 0.68.1 build notes add managers for team bots, so you control who can edit them. Logins go through 1Password: you approve each fill and the bot never sees the password. And auto-review approval cards now describe each action in one sentence and flag irreversible or sensitive changes.

In recent builds. The 0.72 and 0.75 notes from @blankspeaker list sharing an agent into a team workspace, live status on the PR it opened, and connecting a personal Slack account. They also mention a permissions panel for what an agent can use, plus analytics and a review queue for team agents.

Follow this account if you build on Grok Bot - then cross-check the official changelog before you lean on a fresh feature.

Some of that is still marked as not rolled out - like controlling Mac apps or a group chat with teammates around one agent - and the notes warn the stable version may differ.

The rule: share the bot, not your credentials. Every plugin you add to a Team Bot is shared with everyone who talks to it.

Conclusion

One bot is an intern. A roster with a PM, a merge line and an ops bot is an engineering team.

The people building Grok Bot didn't get to 200 parallel agents by writing better prompts. They got there with an org chart: one domain per bot, cloud agents for the code, a board for the truth, a human on the risky merges. Then they gave the team a phone, a store, an inbox and a seat in Slack.

You don't need all ten steps this week. Pick one - hire the Tech Lead, or write the merge line - and do it tomorrow.

The gap between people who chat with bots and people who run them is opening right now.

Source details

Alex · x.comPublished · added here

Original title: Grok Bot just became an engineering team