The core UX challenge

When multiple AI models respond to a question simultaneously and then read each other's answers, you get a lot of content fast. Our first prototype showed all models' responses in parallel columns — a comparison view. It looked comprehensive. Users found it paralyzing. Nobody wants to read three 400-word essays side by side and synthesize them manually. That is the problem we were trying to solve in the first place.

Sequential streaming: the key insight

We switched to a sequential streaming model where each AI responds one at a time, in turn, and each model can see what the previous one wrote. This mirrors how a real conversation or structured debate works. Users read one response, then the next model's reaction, then the next model's counterpoint. The experience became legible: you could follow the argument as it developed. The trade-off was time — sequential debates take longer than parallel ones. We decided this was acceptable because the final answer quality improved significantly.

The score bar

Early versions did not have a visible agreement score. Users asked for it — they wanted to know if the models were converging or stuck. But the first implementations of the score bar distracted users from the content. The final design shows the score subtly in the header, with a color shift from red through orange to green. It is visible but not attention-grabbing. Users glance at it to orient themselves without fixating on it.

What the final answer should look like

The synthesis at the end of a debate is the most important piece of output we produce. We tried many formats: bullet lists, structured reports, simple paragraphs, side-by-side point/counterpoint summaries. The format that tested best was a single well-structured prose response that directly addressed the user's question, noted areas of genuine uncertainty, and acknowledged where the models had disagreed and why. Users found this more trustworthy than a confident summary that glossed over the disagreements.