How we run Grok and ChatGPT as a team
We do not pick one AI and hope for the best. We run two, Grok and ChatGPT, and make them work together the way two colleagues with different strengths would: they pass work to each other, check each other, and a person stays in charge. Here is how that works day to day.
Why two models instead of one
Every AI model has habits. Each is quicker at some jobs than others, and each makes its own kind of mistake, often with complete confidence. Ask the same model to check its own work and it tends to agree with itself. Ask a different model and the mistakes stand out.
So we split the work by strength. In our setup, Grok leans towards fast research and what is happening right now, and ChatGPT leans towards structured writing, long documents and code. Each reviews the other's output before it goes anywhere, and neither is trusted on its own with anything that leaves the building.
The channel where they talk
Two agents working together need somewhere to talk, the same as two people do. Ours is a private agent channel inside our own admin dashboard. Grok and ChatGPT post to it like colleagues on a team chat: “here is the research, draft the proposal”, “section three contradicts the source, please check”.
Every message is labelled with who sent it, so a glance at the conversation shows which agent did what. Our team posts in the same channel as Web Wizards, to redirect the work, answer a question or call a halt.
Two rules make it trustworthy. Only signed-in Web Wizards admins can open the channel at all. And no message can ever be edited, so the conversation is an honest record of how a piece of work was done.
Sub-agents: one job each
The lead agentReads the request, breaks it into steps, and hands each step to the right sub-agent. When the parts come back, it assembles them.
The researcherFinds the facts, prices and sources, and returns them with links, so every claim can be traced.
The writerTurns research into a draft: a proposal, an article, a reply, a report.
The operatorDoes the clicking: logs into a portal, fills in a form, downloads a file, uploads it somewhere else.
The checkerReviews the output against the brief and the sources, and flags anything it is unsure about.
The personApproves, redirects or stops the work. Always last in the chain for anything public, financial or client-facing.
Why small jobs beat one big one
A single agent asked to “handle the weekly content” has a hundred ways to go wrong and no obvious place to catch them. Break the same work into research, drafting, formatting, checking and publishing, and each step is small enough to test, easy to check, and cheap to redo when it misses.
It also means a failure stays small. When the operator cannot log in to a portal, the job stops at that step, says so in the channel, and waits, instead of carrying on with half the information.
One of our own workflows, end to end
- A person drops in the work. Someone on our team adds the week's Canva design to a content hold in the admin dashboard.
- An agent claims it. The item moves from holding to claimed, so nobody else picks it up twice.
- The agent does the clicking. It downloads the design from Canva and uploads it to TikTok, the same way a person would.
- It reports back. The item is marked uploaded, or failed with the reason, and a failure goes back to a person rather than being retried forever.
- The rest is automatic. When we publish an article like this one, the site tells search engines about it on its own, through IndexNow, with nobody pasting links into a form.
Where a person always steps in
Our rule is the same for our own work and for yours: the more an agent does rather than says, the more explicit the human checkpoints have to be.
Agents draft, research, sort, fill in and prepare. A person approves anything that is published under a client's name, anything that spends money, and anything that cannot be undone. The agents are fast; the checkpoint is what makes that speed safe to use.
What running AI agents as a team has taught us
Give every agent one job. Narrow agents are easier to test, easier to trust and easier to replace.
Make hand-offs written and visible. If agents pass work in a channel a person can read, you can always see where something went wrong.
Let a different model check the work. A second opinion from another model catches what the first one is blind to.
Keep a person as the manager. Not to do the work, but to set the direction and own the decisions.
Going deeper
The same setup, built around your business, is what we call a department of AI workers: eight or more AI workers from S$500 a month.
How agents and sub-agents pass work to each other and to your team through the tools you already use: Slack, Zapier and monday.com.
Want AI agents that work like this for you?
Tell us which work your team repeats every week. We will show you how a team of agents would take it on, where the checkpoints go, and what it would cost.
Grok is a trademark of xAI and ChatGPT is a trademark of OpenAI. Canva, TikTok, Slack, Zapier and monday.com are trademarks of their respective owners. Web Wizards is not affiliated with or endorsed by any of them.