Build Your Own AI Team in Grok Bot

You saw my team in the video. This is how you build yours. Steps 1 and 2 take about twenty minutes. No code. (An agent here is an AI teammate with a name, one job and its own conversation, that carries on working after you close your laptop.)

Skip for now: plug-ins, payments, routines, the phone app. Build one agent, watch it work, then add the second.

1. Create your first agent

Click New in the sidebar (or press Ctrl+N, Cmd+N on a Mac), then in New chat choose Create new agent. It lands as "New Agent". Open Bot actions › Edit Profile to fill in its name, job and description.

Paste this into the description and change the words in brackets:

You are [NAME]. Your only job is [ONE JOB, one sentence]. You do not do
anything else. If a task is not this job, say so and hand it back.

Mine is called Scout. Its job is "give me the latest news and updates about AI and marketing, every morning."

Write the job narrow. SpaceXAI's docs warn that a vague job like "General Helper" gives the agent less to go on and makes what it learns harder to reuse. Other agents read this description to work out who a task belongs to, so keep it narrow.

✓ You'll know it worked when you message it one real task and it does that task, without asking you what it is for.

2. Sign it in, once

Guard first. Go to Settings › General › Auto-review and set what needs your approval. Anything that sends, buys, publishes or deletes should stop and ask. Inside a conversation you get Allow once, Always allow and Deny. Use Deny until it has earned better.

Now the login. Your agents share one computer in the cloud, which is what lets them keep going after you shut your laptop. When an agent reaches something needing a login, open Agent Computer from the conversation, take control, do only the blocked step yourself (the password, the code, the CAPTCHA), then tell the agent to continue.

Read this part twice. That sign-in is not private to that one agent. SpaceXAI's docs say to treat any login on that computer as available to every agent you own. So sign in to the accounts you are happy for the whole team to reach, and nothing else.

✓ You'll know it worked when you give that agent the same kind of task again and it does not ask you to log in a second time.

3. Add a manager (only after step 1 has worked twice)

Second agent, different kind of description. This one does no work of its own, it hands work out.

You are [MANAGER NAME]. You do not do tasks yourself. When I give you a task,
check which of my other agents owns that job, hand it to them, and tell me
who has it.

Then you talk to the manager instead of to each agent. Agents can also message each other on their own, but the handoff is easiest to watch in a group chat: New in the sidebar › New chat › select two to six agents (on your phone it is + › New Group Chat). Put the manager and one worker in there and give the manager the task.

✓ You'll know it worked when you can read the manager handing your task to the right agent in that thread, with you not in the middle of it.

4. The rule that saves you

One agent. Watch it do its one job a few times. Then build the second. This is a beta and it acts as you, so a team you have not watched work is a team you cannot trust yet.

✓ You'll know it worked when one agent has run a full week and you can name the single job you want the second one to take.

Do this now (five minutes)

Create one agent. Paste the description from step 1 with your own job in it. Give it one real task from this week and watch what it does. That is your whole first day.

1. Create your first agent

2. Sign it in, once

3. Add a manager (only after step 1 has worked twice)

4. The rule that saves you

Do this now (five minutes)

Join the Front Row.