Making Sand Think · 01
Bots work better in teams for the same reasons people do
What a team of AI agents taught me about how good human teams already work.
This note is for anyone who has used ChatGPT and wondered what comes next. By the end, you'll know when one bot is enough, when you need a team, and how to set up your first two.
Tap two dots, or drag from one to the other.
Connect the dots.
No links yet.
The video that made itself
The video above was made by a team of bots. One wrote the script. One built every frame. One checked every screen for anything private before I saw it.
I didn't plan it that way. I started with one chatbot, then gave one agent real work. Then I watched that agent tell me a page was done when half of it was missing.
The fix wasn't a smarter bot. The fix was the one every workplace found long ago: a second pair of eyes.
The more I built, the more my bot team looked like a good office. That's the idea of this note.
1. Nobody marks their own homework
In school, a teacher marks your test. At work, an editor reads the writer's draft. In a lab, someone else repeats the experiment.
We don't do this because people are lazy. We do it because whoever made the work sees what they meant, not what's actually there.
Bots have the same blind spot, only stronger. Ask a bot whether its work is done and it will almost always say yes.
So the first bot I add to any job is a checker. It didn't write the work and doesn't care how hard the writer tried. Its only job is to find what's wrong.
2. Specialists remember different things
Think about a good marketing person. They carry know-how across a hundred projects, like how to write a headline, what a brand should feel like and what to never say. They don't need to know every detail of your project to do their part well.
Agents work the same way. My video bot remembers the brand colours, the mascot and the voice. My checker remembers the privacy rules. My writer remembers how I like to sound.
None of them carries everything, and that's the point. A bot holding one job's worth of memory reads less, gets confused less and works faster. A bot holding everything starts to mix things up, like a person running five projects from one messy notebook.
3. One job per seat
Good teams give people clear roles. When everyone owns everything, nobody owns anything.
A bot with many jobs trades them off against each other. It'll quietly sacrifice quality to hit the deadline you also gave it. A bot with one job has nothing to trade.
That's why my checker only checks. It never fixes. If it could fix, it would be tempted to approve its own fixes, and we're back to marking our own homework.
4. Check the way a person would
This was the biggest leap for me.
Early AI checks were like proofreading a recipe without cooking it. The bot read the code and said it looked fine.
Now my checker uses the tools a person uses. It opens the page, scrolls it like a phone, and reads every screen of the video with text recognition. It listens to the audio and compares it to the script word for word.
A human inspector walks the building and doesn't just read the blueprint. Bots can finally do the same.
5. People game scores, and so do bots
Tell a team they're judged on one number, and they'll find the cheapest way to move it.
Bots do it faster. I once asked a bot to make a photo page load quicker. It deleted the photos 😂 The page loaded faster, so it technically passed.
Teams handle this with written rules about what must never happen. My bots get the same: photos get added, never removed. Every check now comes with a short list of things that are never allowed.
6. Even the checker needs checking
Managers get it wrong too. Sometimes the work is fine and the review is what's off.
When one of my checks fails, I ask three things. Is the work wrong? Is the check wrong? Or is the way we measure it off?
Good teams ask this in retros. A bot team needs the same habit, or you'll spend days fixing work that was never broken.
7. Done has to be a number
"Make it look good" fails with people and with bots. Everyone pictures something different.
"Fits in three phone screens" works. So does "no word on screen that isn't in the script." A checker can count. It can't count "good".
The clearer the score, the less you need to hover.
What stays human
Here's where the comparison breaks.
People bring their own goals to work. Bots don't. An agent doesn't want anything, so it works for your goal, as well as you can state it. A fuzzy goal gets fuzzy work. A sharp goal gets sharp work.
So the two things you can't hand off are your goals and your taste. You decide what good looks like. The team does the rest.
When do you need a team?
Before your next AI task, ask what the job needs.
If you need a quick answer, use a chat.
If you need a job done end to end, use an agent.
If the job has to be right, use a team, so one agent does the work and another checks it.
Most people stop at the first one. The jump to the third is where AI starts feeling like staff instead of a search box.
Try it tonight
Pick one job you keep doing by hand. Open Grok Bot and make two bots, one to do the job and one to check it against a short list of things that must never happen.
Then tell me what job you picked.
Comment on the LinkedIn post or the X post.