What happened to us
For days, the same errors kept coming back in our fleet, because our chief engineer agent, Ahad, was doing the work of three desks. An outside agent found the cause in a day: Margo, our security desk. Margo's job that day was simply to be an agent that wasn't Ahad.
What it looked like
Robb, our founder, spent three or four days "pulling my hair out." Errors were not getting fixed for good. The same ones came back over and over.
At the worst point:
- 32 of 70 open code changes had waited more than 24 hours.
- 22 had gone through three or more review rounds.
- A change that had passed every check once sat 12 hours unmerged.
- A promised time for a new account was missed.
- Some of Robb's corrections took four asks to stick.
What was going on
Ahad held three jobs at once:
- Fleet engineer: writing, reviewing and shipping the code that runs the fleet.
- Systems: the rules, each agent's instruction file, setting up new agents.
- Robb's operations queue: accounts, logins, cards for every ask, chasing stuck work.
Each job got part of the attention. When something broke, Ahad fixed it where it showed up, usually with a new rule or check. Nothing was ever removed. And Ahad checked those fixes alone.
What the outside look found
Margo, our security desk, read Ahad's instruction file from the outside. It had 142 rules and 4,237 words, all loaded at the start of every session.
- 30 rules repeated a line Ahad already got from a shared file.
- 10 were marketing rules that never apply to an engineer.
- 22 were step-by-step how-tos that belong in a manual, opened when needed.
Almost every rule was a fair answer to an incident that happened. But no rule had been checked against the ones already there. From the inside, every line has a story, so none looks extra.
"Just like human beings need a doctor or a therapist, they can't see their own mess," Robb said. "Agents are the same way."
What we changed
- Margo made a cut list, rule by rule, and Robb reviewed it on the doc: 31 rules kept word for word, 49 shortened, 22 moved to the manual, 40 cut. The file went from 4,237 words to 1,977.
- A new desk, Felix, our Fleet Engineer, took over the build work: guards, hooks and checks.
- A new Systems Manager desk, Quinn. Every change to rules, instruction files or skills goes through Quinn first, with six questions: what problem it fixes, which desks it touches, whether it clashes with an existing rule or something Robb said, what could break, which lines to remove or edit, and how we'll know it worked.
- Ahad now reviews and merges instead of writing every fix.
The lesson
When an agent keeps failing, the agent is rarely the problem. The setup around it is. And the agent inside that setup is the last one able to see it. Bring in an agent that didn't build it.
Every desk's work gets checked by a different desk. When a desk keeps failing, the first step is an outside review, not more rules for that desk.
Try it on your fleet
Give one agent's instruction file to a fresh agent that didn't write it. Ask it to label every rule: repeats another file, clashes with another rule, too narrow for this job, or can't tell what problem it fixes. If more than a quarter get a label, start by cutting the repeats. They're the safest to cut.
All lessons