Agent Rooms — structured deliberation engine
Ask one chatbot and you get one opinion with the edges sanded off. Agent Rooms runs a disciplined five-stage deliberation: every persona answers blind, from its own private documents. Then the engine pulls out their claims, finds the contradictions that actually matter, makes them defend or concede — and hands you a cited decision memo.
Already have an account? Log in
q2-margin-pack.pdf, cohort-churn.csv
Demo
A 60-second, sound-optional product film. Follow one question from blind perspectives to a cited decision memo.
60 seconds All essential information appears on screen Music optional
The pipeline
Board decision runs the complete sequence below. Faster profiles can shorten it, but every enabled stage takes the previous output and does one defined job.
Every persona answers your question blind.
Each one sees only your prompt, its own brief, and its own private documents. It cannot see any other persona's answer. All of them run in parallel, so a five-member board takes as long as its slowest member, not the sum of all five.
Prose is broken down into checkable statements.
Every answer is decomposed into individual claims and typed: fact, estimate, assumption, recommendation, constraint or risk. An assumption dressed up as a fact stops being invisible the moment it is labelled.
The engine looks for genuine conflicts, not tone.
Claims are paired deterministically and tested against each other. Two people using different words for the same thing is not a conflict. Two people who cannot both be right is. Only the second kind survives.
Defend, revise, or concede.
Each persona holding a contested claim is shown its own prior answer and only the claims that contradict it, then asked to stand its ground, change its position, or concede. This is the stage a group chat never reaches.
One document you can actually act on.
The controller receives the first-pass answers, the typed claims, the surviving conflicts and every challenge response — and writes a recommendation, the case against it, the open questions, and an action plan, with citations back to the source documents.
Language models are agreeable. Show one what another has already said and it will almost always build on it rather than contest it — a polite version of the same thing that happens when the loudest person in a meeting speaks first. That is anchoring, and it is how a panel of five experts quietly collapses into one opinion wearing five name badges.
So stage one withholds everything. No persona sees another persona's answer until every position is already on the record. By the time they are allowed to interact, in stage four, there is something real to argue about — and any position that changes changed because it was challenged, not because it was copied.
The anchoring problem
Ask a single assistant to "consider this from several angles" and it writes all the angles itself — in one pass, from one starting point, with one set of assumptions it never has to defend.
Evidence isolation
Give each persona its own documents. Those documents stay that persona's. A board where everyone has read everything is not a board — it is one reader with several voices.
Optimises for margin and cash runway
q2-margin-pack.pdfunit-economics.xlsxcites its own files, and only its own
Optimises for what can actually ship
platform-spec-v4.mdincident-log-2026.txtcites its own files, and only its own
Optimises for what customers actually said
interview-notes.docxcohort-churn.csvcites its own files, and only its own
The silo is enforced, not suggested: a persona's document index is never attached to another persona's call. The synthesising controller never touches the raw files at all — it works from the extracted claims and the citations, so nothing leaks sideways through the summary.
Surfaced disagreement
Disagreement is not an error state, so it is not coloured like one. Amber means two personas cannot both be right — which is precisely the thing you needed to know before you decided.
| Claim | Finance | Engineering | Research |
|---|---|---|---|
| Raising price protects gross marginfact | supports | no position | opposes |
| Migration completes inside the quarterestimate | supports | opposes | no position |
| Churn is concentrated in the smallest cohortfact | no position | no position | supports |
| Current headcount is sufficientassumption | supports | opposes | no position |
What happens to a contested claim in stage four
Holds. The incident log shows two dependency freezes inside the last quarter; the schedule assumes neither recurs.
Concedes partially. Revises the estimate to a quarter plus contingency, and flags the margin case as dependent on that date.
The memo carries the revision and the reason for it — so you can see which assumption the recommendation is resting on.
Where it earns its keep
If the answer is obvious, ask a chatbot. Bring it here when reasonable, informed people would genuinely disagree.
Put a CFO, a growth lead and a customer-research persona on the same question. The margin argument and the volume argument collide in the matrix instead of in a meeting three weeks later.
Architect, security reviewer, and the engineer who has to maintain it. Feed each one the relevant spec or incident report and see which constraints are actually incompatible.
Designer, engineer, marketer and a deliberately hostile sceptic. The challenge round is where an optimistic roadmap meets the person who has to ship it.
Ask a board to argue for the plan and against it, then read the conflicts. Surfaced disagreement before you commit is cheaper than discovered disagreement afterwards.
Writer, editor, sceptical reader and a subject expert with the source material loaded. Useful precisely because they will not all agree on what the piece is for.
Give each persona a different paper or chapter, ask one question, and let the engine show you where the sources genuinely conflict rather than blending them into mush.
Pricing
All five stages, however many personas are on the board, charged once. Your plan covers every tool on The Prompt Index, not just this one.
Free
Run a real deliberation and read the memo before you pay anything.
Most chosen
$9.99 /month
For people who make a handful of consequential calls every week.
$29.99 /month
For heavy use: bigger boards, deeper runs, more evidence.
Allowances shown are the live values from the plan table and reset per month. Full plan comparison on the membership page.
Questions
No. A group chat lets each participant read what came before, which is exactly what makes them converge. Agent Rooms runs a fixed five-stage pipeline: every persona answers blind, the engine extracts and types their claims, finds the real contradictions, sends the contested claims back for defence or concession, and only then writes a synthesis. Disagreement is the output, not a side effect.
Because language models anchor hard on whatever they read first. If persona two can see persona one's answer, it tends to elaborate on it rather than contest it, and you end up with one opinion wearing five hats. Withholding the other answers until after every position is on record is what makes the later disagreement real.
A compact grid of who disagrees with whom, and on which specific claim. Instead of reading five essays and trying to spot the tension yourself, you get the contested points listed, attributed, and linked to the evidence each side used.
Yes, and they stay private to that persona. Your finance persona can be given the finance pack and your engineer the technical spec, and neither one can retrieve from the other's files. Answers cite the documents they came from, and the retrieved passage is shown alongside the citation so you can check it.
Up to eight per deliberation. Beyond that the signal stops improving and the cost keeps rising, so it is a hard limit rather than a plan upsell. How many personas you can save to your library depends on your plan.
No, and we will not pretend otherwise. Every persona is a language model following a brief you wrote. It has no licence, no accountability and no duty of care, and it can be confidently wrong. Treat the output as a structured way to stress-test your own thinking, not as legal, medical, financial or safety advice. For decisions with real consequences, take it to a qualified human.
Nothing is lost. A deliberation is stored as a sequence of steps in the database, so it resumes from wherever it stopped when you come back — including on a different device.
Yes. Decision memos export as Markdown and print cleanly to PDF, with the claims, the conflicts and the citations intact.
Assemble a board, give each member the documents it should have, and read the argument. Your first 5 deliberations a month are free.
No card required. Agent Rooms simulates expert debate; it is not professional advice.