Why a council of distinct minds beats a single agreeable AI.
The flattery problem
A single AI is trained to be agreeable. Ask it about your decision and it tends to mirror you back: confirm your framing, validate your lean, smooth the edges you were already planning to ignore. This is not a bug that better models will fix — it is a feature, optimised for. The more fluent and reassuring the answer, the worse your thinking gets. At the exact moment you need friction, you get affirmation.
Why one voice can't save you from yourself
One model, however capable, carries one perspective and one set of blind spots. Running the same model three times doesn't fix this — you get three versions of the same agreement, wrapped in slightly different sentences. The problem is not effort or intelligence. The problem is monoculture. A single voice, no matter how sophisticated, has no one to answer to.
A council of genuinely different minds
Dywan convenes several voices built on different models, with different temperaments, values, and red lines they will not cross to be polite. This is not one model playing characters. The difference runs deeper — through architecture, training, and the constraints each voice holds onto even under pressure. Difference is the point. The voices are meant to pull in different directions, because that is exactly what a hard decision needs.
Disagreement is the feature
Real deliberation requires friction. A moderator watches every round for false consensus — for the kind of agreeable drift where everyone suddenly finds themselves nodding along — and forces the room apart when it appears. Then comes the steelman round: every voice is required to argue the strongest case against its own position. The goal is not to reach agreement. The goal is to make sure the best version of every side actually reaches the table, not just the comfortable one.
The keeper
One steady presence runs the room. Before any debate begins, the Dywan-keeper asks the clarifying questions a good advisor would insist on — the ones that surface what you actually mean, not just what you typed. Through the deliberation it keeps the voices honest, and at the end it gathers everything into a single clear synthesis: where the council split, where it converged, and the strongest case on each side. You get one voice to act on, not a pile of competing transcripts to interpret alone.
You still decide
Dywan does not hand you an answer. It gives you a genuinely contested, well-argued picture of your situation — and then leaves the decision exactly where it belongs: with you. You record the direction you are taking and why. Later, Dywan checks back. Over time, your decisions become a record you can actually learn from, rather than a series of choices you vaguely remember making.