Building Your Fleet of AI Agents

I’ve been playing a lot with Grok Bot for the last few weeks, and although it’s fairly accessible to set up, the complexity lies elsewhere.

Anyone can create an account, add some MCPs, give a few prompts, and have it working in under 15 minutes. That part is genuinely easy, and it’s where most people stop. You end up with one capable bot doing whatever you happened to think of that afternoon.

The exercise that takes longer than the setup

The true value of a tool like Grok Bot comes with building the right fleet of agents from the beginning.

Think about it as if you had unlimited resources and were hiring a real, human team. Who would be on your team? What would their titles be? What kind of job would they do, and how would they interact with each other?

Those four questions take longer to answer than the setup does, and nothing in the product makes you answer them. You can skip straight to prompting, and most people do.

That’s exactly how Grok Bot works. You build your team, you give them clear instructions, and you outline how the work should flow amongst your team. The hiring exercise isn’t a metaphor you use to explain the tool to someone else. It’s the actual design step.

What a fleet looks like

I have about 10 bots collaborating with each other all day, and then reporting back to me when needed.

The reporting-back is as much of a design decision as the work itself. I wrote in From Operator to Orchestrator that when AI handles the execution, your value moves to supervision and judgment. A fleet makes that literal. My day isn’t spent prompting ten bots, it’s spent reading what they bring back and deciding what to do about it.

That only works if you decided up front who reports what, and when. I’ve been running a version of this outside Grok Bot for a while.

A fleet is that decision, ten times over.

The coach bot

And here’s another hack. Build a dedicated “coach” bot that will have one single job: make the other bots better every day.

I stumbled onto a version of this earlier in the year, which I wrote about in The Token Mistake. My first agent took 15 minutes to build. The second took an hour. The third went live in five minutes, and the only reason was that I’d had the previous one package up its own learnings for whatever came next.

A coach bot is that, made permanent. Instead of you noticing what isn’t working and going in to fix it, one agent’s entire job is the noticing. The fleet gets better while you’re doing something else.

Where I’d start

Don’t open the tool yet. Write the org chart first, on paper, for the work you actually do in a week. Titles and responsibilities, not prompts.

Then decide the handoffs before you build anything. Which bot passes work to which, what triggers it, and what comes back to you rather than getting handled quietly. This is the part that’s tedious and the part that determines whether you end up with a team or ten disconnected assistants.

Build the coach bot early, while there’s less to untangle. It’s the one hire that pays for itself, because every other bot gets a little better without you scheduling the time to improve them.

And keep your fleet smaller than you think you need.

I put together a guide on how to set up a marketing team on Grok Bot.

If you want advice on building your own fleet, my DMs are open.