BetaGoing live early October
DeliboraRequest access

Multi Agent Debates

Where curiosity becomes conversation.

Delibora turns decisions, hypotheses, pitches and research briefs into protocol-governed multi-agent simulations. Distinct AI participants develop, challenge and compare perspectives turn by turn, producing an inspectable transcript and report rather than a single generated answer.

Closed beta. No card, no commitment — we are letting people in a few at a time.

Major models, one room.

Choose a different model for every seat from a curated catalogue of the major providers.

  • Anthropic
  • OpenAI
  • Google Gemini
  • Perplexity
  • Z.ai · GLM
  • Qwen
  • DeepSeek
  • xAI
  • MiniMax
  • Kimi
  • Meta
  • Alibaba Cloud
  • Inception
  • IBM
  • Tencent
  • inclusionAI
  • ByteDance Seed

01Chapter

A room, not a longer prompt.

Delibora stages a structured conversation between several AI participants and persists it turn by turn. Each participant has its own brief and its own private context; a protocol decides who speaks next; exactly one of them speaks per turn.

  • Separate briefs

    Participants are configured individually and snapshotted into the run. What one of them was told is not what the others were told.

  • A protocol, not a free-for-all

    Turn order is a rule of the run, not an accident of who the model felt like answering as. One speaker per turn, every turn.

  • Written down as it happens

    Every turn is persisted when it is made. Close the tab, restart the machine, come back tomorrow — the run is where you left it.

  • You are in the room too

    Queue your own turn and it is taken before the next model turn. Steering a discussion does not mean starting it over.

02Chapter

One answer sounds finished. That is the problem.

Ask a model once and you get a single voice in a single framing, delivered with the same confidence whether it is right or not. There is no dissent in it, no minority position, and nothing you can point at afterwards to see how it was reached.

  • The honest fix is a room

    Several positions, in the open, arguing — and someone taking minutes. That is how decisions get stress-tested everywhere else.

  • Rooms are expensive

    They have to be convened, scheduled and moderated, and the loudest person in them sets the frame long before the best argument does.

  • So convene it in software

    With the structure a real room usually lacks: who speaks when, who hears whom, what each participant was handed, and a transcript that outlives the meeting.

03Chapter

Nine ways to run the room.

Same engine, twelve formats, grouped the way the product groups them. Each one asks a different question about the same idea — argue it, judge it, sell it, test it on an audience, moderate it, or make it compete — and each one names the participants it seats and the artefact it leaves behind.

Evaluate and choose

  • Decision stress-test

    Put a decision you have already half-made in front of two agents briefed to disagree about it.

    They go after the assumptions underneath the decision rather than the wording of it, then converge on a recommendation you can act on.

    Participants
    2 opposing agents
    You get
    Decision memo

    Where it stops: It argues the decision you brought. It will not tell you that you framed the wrong one.

  • Judged debate

    Two answers you cannot choose between, argued as opposing cases and scored by a neutral third seat.

    The arbiter is a participant of its own with its own brief. It scores the exchange rather than joining it.

    Participants
    2 debaters + Arbiter
    You get
    Arbiter’s scored verdict

    Where it stops: It scores the two cases that were actually made. A better third answer nobody argued cannot win.

  • Team Battle

    Two saved Teams take opposite sides of the same question and defend them in a fixed public sequence.

    Every member prepares privately first and only the spokespersons speak in public, so the scorecard measures the cases — construction, evidence, rebuttal, consistency, delivery — and not who talked most.

    Participants
    2 saved Teams + neutral Arbiter
    You get
    Five-dimension scorecard and verdict

    Where it stops: Both sides are assigned their position. A Team winning the round does not make its position the true one.

    Explore Team Battle →
  • Shark Tank

    Take a pitch in front of a panel that opens cold, cross-examines what it heard, and closes without hearing the other closings.

    You say what is being judged — investment, partnership, a purchase, the idea itself, launch readiness — and each Shark scores it independently against that, never seeing another Shark’s private brief.

    Participants
    2–4 Workers
    You get
    Cited scorecard and verdict

    Where it stops: The panel is playing investors. It is not diligence, and it is not an offer.

    Explore Shark Tank →
  • Idea Tournament

    Put ideas head to head in a bracket and get a winner together with the reasoning that killed each loser.

    Every match runs the same frozen protocol with position bias mitigated by construction, so a late round is comparable with an early one.

    Participants
    4–16 ideas, 1–5 judges
    You get
    Champion spec and kill cards

    Where it stops: The bracket ranks the ideas you entered. It cannot tell you about the one you did not.

Explore perspectives

  • Expert panel

    Hand a problem to a small panel of specialists and collect what each viewpoint sees that the others miss.

    The panel deliberately does not vote. You get the insights and the disagreements intact rather than an average of them.

    Participants
    2–6 experts
    You get
    Key insights, no verdict

    Where it stops: There is no ruling at the end. If you need one, run a judged debate instead.

  • 1:1 Human Dialogue

    Two AI personalities talk to each other in a private one-to-one chat, each one believing the other is a person.

    Neither side is told what it is talking to. The chat is masked, there are no phases and no arbiter, and it runs until it hits a limit you set.

    Participants
    2 Worker-backed LLMs
    You get
    Dialogue transcript

    Where it stops: It is the least structured format here. Read it for texture, tone and where a conversation drifts — not for a decision.

  • MultiFormat

    Chain discussions into a workflow: arrange two to five formats so that the result of one step becomes the brief of the next, with checks that route the outcome.

    Each step runs as the format it names and hands the next step a brief built from its result, every item labelled with where it came from; a check between steps can send the workflow down one route or stop it for your review.

    Participants
    2–5 steps, each a format of its own
    You get
    Workflow report with the route taken

    Where it stops: A workflow adds no certainty: each step is as directional as the format it runs, and a check only routes on what the previous step wrote.

Understand audiences

  • Focus Group

    Run a moderated session against an editable discussion guide and get findings that quote what was actually said.

    Opening perspectives first and a wrap-up last, with sections in between; everyone answers a section once, then the moderator either probes the named few or moves on. Airtime is computed on the server.

    Participants
    4–12 Workers
    You get
    Themes and verified quotes

    Where it stops: The panel is generated, not recruited. Nothing here is qualitative research on real participants.

    Explore Focus Group →
  • TribeMind

    Run an idea past a synthetic audience and watch opinion move — or fail to move — round by round.

    Round one is isolated and becomes the baseline. After that each participant hears a fixed set of neighbours and sees only the round before theirs, never the one they are in.

    Participants
    12–50 synthetic participants
    You get
    Metrics and Observer report

    Where it stops: A synthetic audience is a simulation, not a sample. It measures no real population.

    Explore TribeMind →
  • Reach Test

    Screen several versions of a campaign message on a synthetic crowd before launch and see who each one reaches and how it lands.

    Variants are shown to the same synthetic crowd in waves, so each one is compared against the same audience under the same exposure.

    Participants
    1–5 message variants, a synthetic crowd
    You get
    Reach, reactions and a variant ranking

    Where it stops: A synthetic crowd is a simulation, not a sample. Results are directional and measure no real audience.

  • Interview Simulator

    Put your interview guide to a panel of simulated participants before a decision and read a report that cites who said what.

    Each participant answers the whole guide on their own, never seeing the others; a report then counts the closed answers and quotes the open ones back to their source.

    Participants
    2–12 Workers, 1–15 questions
    You get
    Report with cited answers and a question matrix

    Where it stops: Simulated participants are generated, not recruited. Answers are directional and speak for no real customer.

04Chapter

The rules are the product.

What makes a run worth reading is what each participant was and was not allowed to see. Explore how four formats control the exchange: who speaks, what they hear, and how the discussion reaches its result.

Focus Group

How a session is paced

A moderated session has a running order, and in a Focus Group that order is a structure the run enforces rather than a suggestion in a prompt. An opening perspectives section always comes first and a wrap-up always comes last, with one to six topic or stimulus sections in between.

Inside a section the shape is fixed: the moderator asks once, everyone answers once, and then the moderator either probes — putting a follow-up to the named few, not to the room — or moves on. Depth caps the probing, so a section cannot run away with the budget of the ones after it.

Stimulus material is bound to the section that uses it. A card nobody points at is flagged as a warning rather than blocking the run: the session still starts, and you still find out before it does.

Five sections in order: opening perspectives always first, three topic and stimulus sections in the middle, and a wrap-up that is always last, with a bracket under the middle reading one to six topic or stimulus sections.
Opening perspectives first, wrap-up last, one to six sections between.
The turn engine: the moderator opens and asks the question once, everyone answers once in a response sweep, then the moderator either probes — a reaction sweep of only the named few, up to two probe rounds per section on standard depth — or moves on to the next section, ending at completed.
Ask once, sweep once, then probe the named few or move on.
Guide sections on the left bound to stimulus cards on the right: section 3 points at card 1, section 4 points at card 2, a topic section uses no card, and card 3 is not bound to any section — an unbound card is a warning, never a blocker.
Every stimulus card belongs to a section, or it is flagged.

Team Battle

Two Teams. One motion. A reasoned verdict.

Give two saved Teams opposing positions on a concrete decision — for example, whether to launch a weekly newsletter before expanding paid ads. Both sides receive the same facts and evidence, while each Team keeps its own members, models and working style.

Each Team prepares privately before its spokesperson opens the public debate. Questions, answers and rebuttals put the two cases under pressure; closing statements give each side a final chance to make its case.

A neutral Arbiter, who sits on neither Team, scores five dimensions and delivers a verdict with reasons. The public transcript and scorecard show where each argument held up. A win judges the cases presented; it does not prove that a proposal will succeed.

Two Teams share one motion and the same evidence. Members prepare privately; one spokesperson per Team debates publicly, and a neutral Arbiter weighs the exchange.
Shared facts, opposing positions, and a neutral Arbiter.
Five stages: Prepare in three private huddle blocks, Open with both cases, Challenge through cross-examination and rebuttals, Close, then Judge with five scores and a verdict.
Private preparation leads to a structured public debate and a scored verdict.

Shark Tank Pitch

What a participant is handed

A panel that has read everything before it speaks is a panel that agrees with itself. The Shark Tank pitch takes the transcript away from the phases where it would do damage.

An opening sees no turns at all — just your pitch and whatever else you filled in — so the first reactions are genuinely independent. Cross-examination is the phase that gets the room: every earlier turn except the closings. And a closing sees your pitch and the cross-examination, but never another closing, so the last word is not simply the previous last word agreed with.

A Shark is also never shown another Shark’s private role text. What makes one of them sceptical about unit economics stays theirs.

Three phases on the left — opening (sees no turns at all), cross-examination (sees every earlier non-closing turn) and closing (sees no other closing) — with arrows to what each is handed on the right, and crossed-out dashed arrows to the turns it is not allowed to read.
Each phase is handed a different slice of the room.

TribeMind

Who hears whom

A synthetic audience that all reads the same feed is not an audience, it is one opinion repeated. So in TribeMind nobody hears the crowd. Each participant is given a fixed set of neighbours and hears those, in every round, and nobody else.

The connection runs one way: participant 1 hearing someone does not mean that person hears them back. That asymmetry is what lets an opinion travel through part of the audience instead of arriving everywhere at once.

Rounds are sealed the same way. Round one is isolated — the stimulus and nothing else, not even the knowledge that anyone else is answering — and it is the baseline every later round is measured against. From round two on, a participant sees what their neighbours said in the round before, never what those same neighbours are saying in the round they are in.

A circle of twelve participants highlights one listener and the six neighbours whose contributions they receive. Every participant has their own fixed set; hearing someone does not mean they hear you back.
Every participant hears a fixed few — and not the same few.
Two panels. Round 1, isolated: the stimulus cards and nothing else — no neighbours, no totals, no idea that anyone else is answering. Round 2, social: the stimulus cards plus what their six neighbours said in Round 1, while what those neighbours are saying in Round 2 is not shown to them.
A round sees the round before it, never the one it is in.

05Chapter

What you take away.

A run is not finished when the last participant stops talking. It is finished when you have something you can hand to somebody else.

  • The transcript

    Every turn, in order, as it was persisted. It is the evidence behind the summary, and it is never gated — every plan keeps it.

  • The report

    A PDF written for whoever was not in the room: what was decided, what was contested, and what the run was actually asked. Also on every plan.

  • Your own evidence

    Upload documents and the participants argue against what is in them. Excerpts are bounded on purpose — a whole file in a prompt is not grounding, it is noise.

  • The model that answered

    Choose which model sits in which seat. Reports name the model that actually produced a turn, not the one the seat was configured with.

06Chapter

What this is not.

Delibora simulates a room. It does not sample a population, and we would rather say so on the landing page than in a footnote after you have paid.

  • Participants are language models playing a brief. A TribeMind study is not a survey and a Focus Group is not recruitment — no number either one produces is a measurement of real people.
  • What it is good for is rehearsal: surfacing the objections, framings and failure modes of an idea before you spend real time and real people on them.
  • Every run shows its work. If a conclusion looks wrong, the transcript is right there, and you can read exactly what the participant who said it had been handed.

07Private beta

Delibora is in private beta.

We are opening it a few accounts at a time. Tell us what you would run and we will come back to you when there is room — three questions, about a minute.

  • Free while the beta runs — no card, no commitment.
  • We read every answer; they decide what we build next.
  • One email when there is room for you. No newsletter.

Step 1 of 3

Your field

  1. Step 1, current step
  2. Step 2
  3. Step 3
What industry or niche are you in?

Choose the field that best describes your work.

By joining you agree to receive infrequent product updates. Unsubscribe anytime. We handle your address as set out in the Privacy Policy. Questions? hello@delibora.com