In shortYour question opens a group conversation. The models take turns, each one reading everything said so far, agreeing, disagreeing and refining. A live agreement score tracks how close they are, and it runs until you stop it.
A shared conversation, not a pollAlethe doesn't ask each model in isolation and average the answers. Every model sees the full transcript before it speaks, so it can build on a good point, correct a wrong one, or push back on another model by name. The conversation is the product, not a set of parallel monologues.
Casual or substantive: it adaptsIf your question is small talk or a simple fact, the models just answer naturally, like a normal chat. If it's a real question(an analysis, an opinion, a technical problem)they dig in, take positions and challenge each other. You don't switch anything; they read the room.
It never forces a conclusionThere's no judge and no automatic verdict. The conversation doesn't stop itself once the models 'agree enough', it keeps going until you stop it or you reach your plan's token limit. Convergence isn't announced; you watch it happen, both in what the models say and in the agreement score.
The agreement scoreAs the models talk, Alethe measures how semantically close their latest messages are (using vector embeddings) and turns it into a live 0–100% agreement score, with a per-model breakdown. High means they're closing in on the same answer; low means the question is genuinely open, a signal to read both sides rather than trust a single line.
You stay in controlDrop a message in at any moment to add a constraint or change direction; the models pick it up on their next turn. You can add a new model to a running conversation, or nudge who speaks next. Stop whenever you have what you need, and reopen the conversation later exactly where it left off.