Six frontier AIs debate.
One synthesizer answers.
Ask one hard question. Six frontier AIs argue it for up to ten rounds. You walk away with one cited report — verdict, tensions, sources — that's yours to forward or revisit. The report is the product, not the transcript.
Two tiers, real models.
Two frontier models debate. Perplexity web research is always on — it's the third leg of the value prop, feeding both agents shared, current facts.
GPT-5, Gemini 3.1 Pro, Claude Sonnet 4.5, Grok 4, DeepSeek V3, Perplexity Sonar. Web research toggleable. File uploads allowed. Deeper report.
No personas. Real difference.
We do not assign personas. Nobody is told to be "the skeptic" or "the optimist." Every agent gets the same system prompt.
The disagreement is emergent. Different training data, different RLHF, different reasoning styles — put them on the same ambiguous question and they naturally diverge. That's the entire thesis of the product.
Four constraints, nothing else.
You pick Decide, Stress-test, Research, or Compare. Every agent sees the same system prompt for that intent — e.g. Stress-test tells them to identify failure modes the others missed and reject vague generalities.
Before the first turn, Perplexity runs a neutral search on your question and injects a summary + sources into every agent's context. Nobody argues from stale training data.
Each turn is capped at ~180 words. This kills waffle, forces specificity, and keeps the round-trip fast.
Agents are instructed to cite prior turns by agent name when they disagree, and never to repeat themselves. Agreement is allowed — but it has to cost words, so it only happens when they really do agree.
Round-robin, full context.
Default is 8 rounds. Agents take turns in a fixed rotation. Each turn sees the full transcript so far — every prior agent's argument, verbatim.
Round 1 — Agent A opens. Round 2 — Agent B reads A, responds, may push back. Round 3 — Agent A reads B's pushback, revises or defends. Round 4 — Agent B reads Rounds 1–3, sharpens. ... Round 8 — Final word. → Synthesis Engine reads the whole transcript.
Not kumbaya. Not fighting for sport.
- →Frontier models from different labs really do disagree on ambiguous questions. That's the well-documented behavior — not a trick we engineer.
- →The intent prompt (especially Stress-test) explicitly rewards finding holes in the prior turn.
- →The word budget makes empty agreement expensive, so it only happens when it's real.
- →We do not assign personas. Nobody is told to be the skeptic or the optimist.
- →The Decide prompt says explicitly: update your view only when evidence warrants.
- →The synthesizer at the end rewards genuine convergence, so agents aren't punished for agreeing.
A separate model reads the transcript.
After the last round, Gemini 3.1 Pro reads the entire transcript with a separate synthesis prompt. It doesn't add new arguments — it distills what happened.
- · Verdict + confidence level
- · Points of agreement
- · Tensions, sorted strongest disagreement first
- · Cited sources
- · Minority Report — the strongest dissenting view, preserved
- · Executive summary
- · Risks with severity + mitigation
- · Unknowns that would sharpen the decision
- · Decision framework — the branching questions
- · Scenarios (base / best / worst)
- · Action plan (week 1, month 1, quarter 1)
The Minority Report is deliberate: when one agent stakes out a strong contrarian view, the synthesizer preserves it instead of averaging it away. Consensus is not always the truth.
Debate Mode.
Today the mechanism is emergent disagreement — models disagree because they're different. A future Debate Mode will assign opposing sides, so you get structureddisagreement on questions that benefit from it. Both are useful; they answer different questions.