Proposer
claude-sonnet-4-5 Multi-model adversarial deliberation
The room where AI argues with itself.
Routers optimize cost. Agents execute tasks. Think Council stress-tests decisions — by making four AI personas debate yours, then synthesizing the recommendation with the dissent attached.
~400 on the waitlist invites Fridays no card
The council
Claude GPT Gemini Llama
TOPIC Turn 1 — opening arguments
Challenger
gpt-5 Steelmanner
gemini-2.5-pro Pre-Mortem
llama-3.3-70bSynthesis 0%
Recommendation. Use Auth0 now, plan to migrate at 5k MAU. Time savings justify the cost at your current stage.
The cast · always these four
Four roles. Forced.
Hue is the only thing that varies.
Every deliberation runs the same shape. The personas read each other across turns and rebut — they don't vote in parallel. That's the difference.
Proposer
ForBuilds the strongest case.
Frames the upside, names the trade-offs, anchors the decision.
Challenger
AgainstAttacks every assumption.
Surfaces blind spots, hidden costs, lock-in. "What am I not seeing?"
Steelmanner
BothArgues the best version.
No straw men. Reconciles where reconcilable, names the real fork where not.
Pre-Mortem
FutureWrites the obituary.
Narrates how this fails 18 months from now. Validation gates fall out of it.
Why a council, not a router
Routers optimize cost.
Agents execute tasks.
Councils stress-test decisions.
One opinion, fast.
You ask, it answers, you ship. Great for low-stakes calls. Bad for the ones that compound.
No disagreement
Disagreement, averaged out.
Routes to the cheapest model, or votes across many. Optimizes cost or consensus — both wash out the dissent that mattered.
Disagreement discarded
Disagreement, preserved + cited.
Four personas argue across turns. The synthesis tells you what to do — and shows you exactly which objection it overruled to get there.
Decision quality
Decisions that earn a council
Built for the calls that compound.
"Auth0, or roll our own?"
Council surfaces lock-in, cost-curve break points, migration pain.
"Postgres or DynamoDB for events?"
Pre-Mortem narrates the migration story you'd rather not write.
"Ship the AI feature this quarter?"
Steelmanner reconciles velocity vs. retention risk into a defensible call.
"Senior IC, or two mid-levels?"
Challenger names the second-order effects. You bring it to the room.
From the founder
"I kept catching myself trusting whichever model answered first. The fix wasn't a smarter model — it was making them argue."
Frequently disputed
Questions worth answering.
Won't the four models just agree with each other?
They're prompted into adversarial roles with different cognitive stances. Agreement is rare — and when it happens, that's a signal, not noise.
How is this different from a multi-agent framework?
Agents execute. Councils deliberate. The output is a single recommendation with the dissent attached, not a chain of tool calls.
What does it cost?
Free: 3 deliberations/day with 2 models. Pro at $29/mo: unlimited, all four models, custom personas, BYOK, full history.
When do invites go out?
Weekly waves. We rate-shape so the streaming stays fast for everyone in the council.
Is the library actually open source?
Yes. MIT licensed on npm and GitHub. The hosted product is built on top.
Waitlist · invites Fridays
Show me the dissent,
not just the answer.
Join the waitlist. We'll send your invite the next Friday.
~400 ahead of you · no card · no sales call