I built this, and here’s what broke.
I run my business with an AI agent team on Claude. Agents draft, research, edit and build. A human decides what ships.
The guardrails are in the code.
Every limit stops the run in code, so no human has to remember it.
One lane acts alone
Everything else waits for a human to say yes.
Spending has a ceiling
A guard pauses paid video renders at 70% of the monthly credits. Only a human can approve more.
Every agent is labeled
When an agent appears in a video, it says it’s AI. Every time.
One season, broken out.
Here’s what the humans caught in one season of the show.
An agent had the wrong face for two episodes
It went unnoticed until a human watched the cut.
Eight episodes were uploaded six decibels too quiet
A check caught it minutes before the posts went out.
A film ended on the wrong shot
A filename was trusted instead of opened. Now the file gets opened.
Making it was cheap. Deciding, checking and throwing work away is the expensive part. That’s the human in the loop.
In the Loop
Essays on people and AI agents working together. Sources linked, no hype.