Most people scaling up an AI assistant make the model bigger. I went the other way: I split mine into thirteen.
Each one has a narrow job and its own memory. Research digs and verifies. Content writes in a voice, not a template. Finance tracks budgets and won't touch a metaphor. Strategy thinks in quarters, not minutes. Critic's entire function is to disagree with the others — devil's advocate as a permanent role, not an occasional prompt. There's one that does nothing but psychology — sessions, reflection, the parts of running a life that aren't tasks.
Why not one model that does all of it? Because specialization isn't just division of labor — it's division of judgment. A general-purpose agent trying to be a copywriter, an accountant, and a therapist at once ends up flattening all three into the same cautious, generic voice. Split them, and each one can actually commit to a position instead of hedging across every possible role it might need to play next.
The friction is a feature. When Critic pushes back on something Strategy proposed, I don't see that as two subsystems out of sync — I see it as the check that a single monolithic model can't give itself, because it's the same weights arguing with itself in one context window. Thirteen separate agents can actually disagree, on the record, and I get to see the disagreement instead of a smoothed-over average.
It's slower to build this way. Coordinating thirteen agents is real infrastructure — memory that doesn't leak between them, a way for them to hand off work, a way for me to know who said what. But the alternative — one god-model doing everything adequately — was never actually what I wanted. I wanted specialists who could be wrong loudly, in their own voice, instead of one generalist quietly hedging.
Top comments (0)