The Factory With the Lights Off (When Nobody Can See the Agents)
The Problem
You told everyone to build, and they did. Hundreds of skills and agents in two months. That's what you wanted. Then something goes wrong, and Legal asks the obvious question: how many agents do we have? Nobody knows how many are running, what they read, who owns them or whether anyone uses them. When you finally pull the logs, it's worse than you imagined. Dozens with no owner, because the people who built them changed jobs. Dozens of near-copies of the same weekly report helper. Several versions of the same policy, giving customers different answers. And the number that matters most: plenty of people use what they built themselves, and hardly anyone uses what a colleague built. That's twenty one-person companies in every department, all with the lights off.
The Principle
Turn the lights on. Every agent and shared skill goes on a screen, with an owner, a backup and a usage count. After a scare, the instinct is to freeze everything. Don't. The builders won't stop building. They'll just stop telling anyone. Make it visible instead, then ask the questions you'd ask about dashboards: who uses it, and what does it drive? Build the registry from the platform's own logs, not from forms. Give every orphaned agent thirty days to be adopted or switched off. Merge the duplicates into one shared skill with an owner, and let whoever thinks theirs is better make the shared one better. Then watch two numbers: people using what they built, and people using what someone else built. The gap between them is your next bottleneck.
5 Steps to Take Action
- 1
Count from the logs
How many agents people think you have, against how many the logs say. Then the share with a named owner and a backup.
- 2
Put them on a screen
What each one does, its owner, what data it reads, what it can do and who uses it, where everyone can see.
- 3
Adopt or switch off
Thirty days for every orphan. Most won't be missed.
- 4
Merge the duplicates
One shared skill per job, with an owner who signs off on every line. Anyone who thinks theirs is better gets to make the shared one better.
- 5
Watch the two numbers
Using what they built against using what others built. A high first number and a low second one means one-person companies and no team.
💡 Inspired By
Chapter 9 of The Second Factory, "Twenty One-Person Companies." And Sarah Chen's forty-seven dashboards in The Auction Block: she didn't ban them, she made them visible, asked who used each one, and killed the rest.
This is what Analytics Pitfalls & Principles is about: people helping people.
In Practice
The morning after an agent emailed hundreds of customers about parts ClassicMotors didn't have yet, the general counsel and the head of IT were waiting outside the CEO's office. "We need to stop every agent in the company," the general counsel said. "Today." "How many agents would you be freezing?" Nora asked. Nobody knew. She asked for until Friday to turn the lights on. Jin Park put it plainly on the catwalk. "We spent two years putting lights on this floor, and then we built a second factory right next to it and left the lights off." On Friday, the new screens went up. Agents running: 612. With no owner: 74. Talking to customers: 41, nineteen of them reading from data rated under three stars. Thirty-eight versions of a weekly report helper. And two lines Nora read over and over. People using a skill or agent they built themselves: 23%. People using one someone else built: 4%. "I told everyone to become a one-person company," Nora said. "You told everyone to climb the ladder," Jin said. "They did. They just all stopped at rung three."
