Grok Bot: How I'd Run a Design Studio with 15 AI Agents

The video above is a thought experiment, so let me be honest about that first. I haven’t opened this studio. I run a 100-person design studio inside a very large company, which is a different animal. But I’ve been beta testing Grok Bot for a few months now, and somewhere along the way I started sketching the studio I would build if I started from zero tomorrow. This is that sketch.

The rule that does all the work

Every studio I’ve ever worked in splits roughly in half. One half is the reason people became designers. The other half is quotes, invoices, chasing payments, deploying sites, backing them up, writing the LinkedIn post nobody wants to write, keeping the SEO alive, and the weekly argument about who is free next Tuesday. The first half is why we exist. The second half is why we’re tired.

So the only real decision in this whole setup is a rule. If a task is repetitive, measurable and draining, a bot does it. If it needs taste, judgment or a face in front of a client, a person does it. Everything else in the video is just that rule applied fifteen times.

Why one bot talks to me, and fourteen don’t

The mistake I see people make with agents is building a dozen of them and then spending their day managing a dozen chat windows. That’s not delegation, that’s a new inbox.

In this studio I talk to one agent. The Chief of Staff holds the studio’s memory, reads and writes Monday.com, and every morning at seven sends me a brief: who is overloaded, what is due, which invoices are late, what needs my approval. When a new project is signed, it reads the budget per role, looks at who has the skills and the hours, and proposes two or three teams with the margin each one leaves. I pick one. It writes the allocations. I never see the other fourteen bots unless I want to; they report to it, not to me.

Grok Bot makes this shape natural because bots can delegate to other bots inside a shared thread. That one feature is the difference between a team and a pile of assistants.

The two bots I’d build first

Most of the roster is obvious once you’ve seen it. Scout finds clients. Proposal writes quotes. Finance invoices through Stripe and waits for a yes. Builder deploys on Vercel and keeps everything on GitHub. Fine. The two I’d spend real time on are the ones that protect the design.

Brand Keeper reads each client’s design system straight from Figma, tokenizes it with Tokens Studio, compiles it with Style Dictionary and keeps a Storybook alive with every component and its states. The point isn’t the tooling. The point is that every other bot, when it builds a page or spins up an A/B variant, pulls from the client’s real components instead of inventing a button. A designer decides what the brand is. The bot makes sure nobody drifts from it, including at 3am when nobody is watching.

Synthetic Lab runs a panel of synthetic personas for each client, calibrated on real data from the CRM, the analytics and the interviews we’ve actually done. Before a concept reaches the client, it goes through ten rounds with that panel in an afternoon. We arrive at real user testing with two strong variants instead of six weak ones. I’d run it on MiroFish, which is open source and lives on the bot’s own machine, so the client data that shapes the personas never leaves the studio. And a researcher still owns the personas, decides whether to trust the signal, and never presents a synthetic result as research with real people. That line matters more than the speed.

What I’d never hand over

Creative direction. The first conversation with a client. Any decision about money. Any judgment about a person. The bots prepare, propose and wait. Humans decide, present and take responsibility. If you take one thing from this piece, make it the list of what stays human, because that list is the actual design of the studio. The bots are just the plumbing.

What you should care about

The beta is honest about its gaps, but you can work with it. There is no fine-grained permission model yet and no proper audit trail, so you build those yourself: one account per bot with minimal rights, three levels of action (do it and log it, prepare and wait for a yes, never), and a log board in Monday that records every approved action and who approved it. And you follow SpaceXAI’s own advice, which I’d frame on the wall: do the task once by hand, then automate it. Automation multiplies whatever the first run did, including the mistakes.

If you build a version of this, I’d love to see where it breaks.