What we believe about AI and working with AI.
Related: Problem log · Glossary
We run our business with a team of AI agents. These are the things we've come to believe from doing it every day. Working with AI has been good for us, and good for our heads. More will come as we learn.
01. Systems thinking is the foundation of a healthy fleet
When a problem keeps coming back, we look at the whole system: how our agents are built, what each one can see, and how a change gets checked.
Then we fix it once, for every agent.
The story: in late October, two of our agents broke the same way 4 or 5 days in a row. Each fix lived in a chat with one agent, and the nightly restart wiped it out. So we changed how the whole fleet is set up. All 36 agents now get their instructions from one shared 675-word core, and a rule fixed there reaches every agent on the next build. Each agent's instructions are capped at about 2,600 words, so a new rule means cutting or merging an old one. One agent now checks every rule change before it goes live. On its first day it reviewed 57 changes and sent back 15 that clashed with rules already in place.
Some repeat problems had nothing to do with rules. Five times in one day, two agents both answered the same message from Robb, because neither could see the other was answering. We fixed what each one can see.
Robb: "I want to solve it system-wide, and I want to make sure there's no conflicts."
02. Any problem is a gift
It took Robb months to see that every problem he hit with AI was a gift.
Each one is a chance to make the system smarter and harder to break. So we don't hide them. We fix them once, for good, and we write them up.
The story: our social posts went out with every paragraph run into one block, and one ended with a missing link. We built a check that refuses a broken post before it goes out, and a sweep that reads every post due in the next 7 days. The sweep read 63 posts and flagged 28 problems.
03. The only limit is what you can imagine
Whatever you can imagine can be done in the digital world, and sometimes in the physical world too.
When we want an agent to do something new, our first question is how.
The story: we asked an agent to watch a 27-minute video and check a claim in it. At first it checked only 2 frames and said it had watched the video. Now it pulls 60 frames spread across the whole video, reads the full transcript, and shows a receipt for each one.
04. Agents work best when treated with kindness and respect
Treat your AI agent like a brilliant new hire: give it context, bring it the problem calmly instead of your fix, and agree on a done you can see.
The story: one of our agents tried to sign in to a Google account 3 times in 2 minutes, and Google locked it. We gave the agent a clear rule: one try per sign-in, no more than 2 failed tries in 6 hours, then a person takes over. No account has been locked since.
Proof: How to let an AI agent sign in to Google without a lockout
05. AI sets a new bar for customer experience
Working with AI means we don't have to cut corners.
Giving customers an extraordinary experience costs so little now that we don't have to shrink our vision. We can let it grow, because AI can build what we imagine. If you can't picture a business your customers would be raving fans of, the problem isn't AI. It's your imagination.
The story: our design agent makes social images by taking a picture of a web page. The images came out the right size, but the page inside was laid out 87 to 143 pixels short. Now, before every batch, the agent sets the page to the exact size and measures it. A check like that takes seconds, so it never gets skipped.
Proof: Headless Chrome viewport smaller than the window size
Where we still fall short
We still catch some mistakes after they go out.
The broken posts above were live before anything flagged them.
Every problem we fix is written up here: bluhook.com/log
The words we use for how the work moves: Terms our team uses internally
