I had fourteen AI agents, and I couldn’t decide which ones to cut.
I’d built each one for a reason. Research, writing, editing, coaching, Obsidian vault management, code, distribution, travel. Every agent had a brief and a defined role. The system still felt oversized. Two more AI models had joined the workflow. Several roles overlapped. I kept circling the same vague goal: simplify the system.
Then I pulled the usage data.
Giulia, my vault manager, had been dispatched 17 times. Leo, the developer, 12. The three editorial agents had been dispatched 2, 3, and 6 times. Ren, who handled distribution, had 7.
The counts didn’t make the decision for me. Three coaches showed zero dispatches because I spoke to them directly rather than sending them background tasks. Removing them on usage alone would have mistaken the measurement for the work. The travel agent showed the same zero, and that one meant what it said. I’d built it and never used it. The editorial roles were a third case. Their functions were already covered by reusable skills (written procedures any of the models can run) and cross-model review. Ren’s wasn’t.
That was the unresolved part. To make sense of any count, I’d had to ask what job sat behind it and what else could carry it. I wasn’t trying to make the roster smaller. I was deciding which jobs still needed a named agent.
Once the decision had a name, the options were easy to list. Keep all 14. Remove the one unused travel agent. Retire the editorial pipeline. Or remove distribution too and handle it ad hoc.
I retired the three editorial agents and the travel agent. Research, drafting, and editing moved to skills that could run across Codex, Claude, and Gemini. Ren stayed because distribution still had a distinct job and no replacement. The roster went from 14 to 10.
The system felt cleaner, but not because ten is a better number than fourteen. The vague simplification problem had become a boundary decision: named agent or reusable skill. Once every role had to sit on one side of that line, the overlap was easier to see.
The cut had a cost. I lost a dedicated editorial gate and now have to run the review skills deliberately. The complexity didn’t vanish, but some of it turned into a trade-off I could name and accept.
That’s the diagnostic I use now. When something in a personal system feels complicated, I ask what hasn’t been decided: the goal, the boundary, or the trade-off. Sometimes the answer is nothing. Some problems are genuinely intricate. But if I can’t explain what the system is for, where one role ends, or what cost I’ve accepted, the complexity is often carrying a decision I haven’t made yet.
