Most People Are Still Using One Chat Box. SpaceX Engineers Have Already Raised 14 Coworkers

An engineer at SpaceXAI filmed a 45-minute video and laid out, all at once, the 14 Grok Bot agents the company uses internally. From the Chief of Staff that manages schedules, to the Figma bot that turns a single keyframe into a complete design, to Dr. Eggbot, which specifically audits the other bots, each one comes with a timestamp.

Here's the list in the order it appears in the video:

0% → 1:23, Chief of Staff bot, manages daily tasks
15% → 6:31, PM, designer, and engineer bots collaborating
30% → 12:43, Figma bot, produces a complete design from one keyframe
50% → 17:59, Dr. Eggbot, builds bots and audits bots
65% → 25:22, engineering lead bot, manages an entire group of agents
80% → 32:34, social bot, turns feedback on X into engineering tasks
100% → 38:59, turns skills into routines, making agents more autonomous

The comments were direct. Someone wrote “45 minutes pleasure,” someone else said “45 minutes alpha,” and another person wanted to save it for the weekend. When a 45-minute workflow video gets that kind of response, it usually means there's no fluff in it.

But to understand these 14 bots, you first need to know something else.

On October 6, the shell and the brain split

On the evening of October 6, Musk posted on X: going forward, SpaceX will choose the best backend model for each task, including Claude Opus 5.5, Midjourney, and Suno, with the standard being “most likely to produce the best result.” A few hours later, Lauren from the Grok Bot team added: all bots are powered by Opus 5.5 and can also spin up cloud agents running Cursor models. 9to5Mac reported on this on October 7.

Using a competitor's model has a simple reason. Artificial Analysis tested both companies on Terminal-Bench using the same harness, and the gap was nearly 3x. There is a footnote to the data: the test was on Grok 4.6, not 4.7, which was released in September; the two measurements were a month apart. But Terminal-Bench measures precisely whether you can hand off an entire job and have it completed, not whether it can answer a single question correctly.

Morlex made this clear in the tutorial: over the past three years, every company's agent was sold as an all-in-one package, with the shell (computer, login state, memory, routines) and the brain made by the same company. After October 7, SpaceX kept the shell and swapped the brain for Anthropic's.

If you want to copy it, these are the most important steps

Most of the ten-step tutorial is common sense; here are a few easily overlooked points.

Write a job for the bot, not a prompt. A prompt is a one-off request; a bot is a role that accumulates memory and claims a domain. Name it by role: Inbox Manager, Research Lead, Sales Outbound. Spell out what it owns, what counts as good, and where it must stop and ask you.

Login walls are handled by handoff, not passwords. In real companies, most software has no API and no MCP server. The bot clicks through to the login page in its own cloud browser, hands you the screen, you log in and click done, and it continues using the same session. The bot gets a session, not a password. Any tool that asks you to paste a password into a chat box has the wrong approach.

Record the screen once, and turn it into a routine. Walk it through the task once; it saves the steps and does it itself next time. Choose weekly work that spans two or more apps and has stable steps. State the trigger conditions in plain language; no need to draw a flowchart.

A bot can spin up Cursor agents. This is the quieter half of the announcement. SpaceX acquired Cursor in August this year for $60 billion. A coding bot therefore becomes a manager: Opus does planning and context, Cursor's agents work in parallel, and the bot comes back to accept the result.

In group chat, give goals, not task lists. A task list means you've already done the decomposition. Give goals and let them divide the work themselves. The example SpaceXAI gives is an engineering bot reproducing a bug, opening a ticket, and passing it to another bot to debug.

Set the approval line by reversibility, not task size. If the bot can undo it, let it finish on its own. If it touches the outside world, moves money, or cannot be undone, stop and wait for you. The desired state is: 36 drafts queued, 0 sent.

Two questions that still have no answers

A commenter named Danny asked: how do you know it's actually using that model? If it doesn't tell you itself, users have no way to verify it.

There's also a more practical one: where does the data go? When the bot thinks with Opus, what it reads is sent to Anthropic. Tom asked about opt-out under Musk's post and got no reply. SpaceXAI also hasn't published sharing terms. The only choice you have is what the bot touches; you can't choose where the model runs.

Cost is also disputed. A commenter named Guillermo said token consumption is too fast and doesn't recommend using it. In Grok Bot you don't pay per token; the cost is included in SuperGrok or a paid Cursor plan. Adam Hincu asked whether Opus-powered tasks would burn through plan quotas faster, and SpaceXAI didn't answer. Comparing API prices, for the same small task—30K input, 8K output—Opus is about 28 cents and Grok about 11 cents. With very long context it flips: Grok's price doubles above 200K tokens, while Opus stays the same.

Also, Midjourney and Suno aren't live yet, with no timeline. One user's bot said on the first day that Suno was unavailable. What you can rely on is Opus 5.5; wait for the rest.

The shell is the product

Morlex's judgment can serve as a reference point: for three years everyone has been asking which model is best, and this week the most aggressive model vendor changed the question to which shell is best, then stuffed its competitor's brain inside.

The shell holds the computer, login state, memory, and routines, while the brain underneath can be swapped at any time. What you need to decide is what to hand over, where the bot's permissions stop, and which data it can touch. Write these three things into the charter.

You can watch the video as an operations manual: 45 minutes, 14 roles. After watching, decide whether you want to raise one yourself first.