A structured multi-agent debate platform for decisions, pitches, research, and ideas. Give several AI participants the same question, compare their arguments, and leave with an inspectable result.
The platform's debate workflows are informed by published multi-agent research. The references below explain the methods and tradeoffs behind structured debate, judging, and collaboration on arXiv, ACL Anthology, or NeurIPS.
New to the field? Read our guide on when to use multi-agent debate vs self-consistency.
Du, Li, Torralba, Tenenbaum, Mordatch
Foundational result: agents critiquing each other across rounds converge on more factual, better-reasoned answers.
Liang, He, Jiao, Wang, Wang, Wang, Yang, Shi, Tu
Introduces the Degeneration-of-Thought problem and shows multi-agent debate unlocks divergent thinking where self-reflection stalls.
Feng, Zhao, Lyu, Li, Tu, Wang
Introduces the per-dimension arbiter sweep that informs the platform's truth-seeking verdict scoring.
Chan, Chen, Yu, Lu, Sun, Liu
Demonstrates that multi-agent debate panels evaluate generated text more reliably than single-judge baselines.
Khan, Hughes, Valentine, Ruis, Sachan, Radhakrishnan, Bowman, Perez
Strong empirical evidence that debate makes weaker judges reliably select truthful answers from stronger debaters.
Hu, Tan, Wang, Qu, Chen
Formalizes debate among LLM judges and adds adaptive stability detection so debates stop when consensus stabilizes — improving accuracy over majority vote.
Wu, Bansal, Zhang, Wu, Li, Zhu, Wang, Saied, Awadallah, Awadalla, Wang
Shows that role-specialized agent groups orchestrated through structured conversation consistently outperform monolithic prompts on complex tasks.
Shinn, Cassano, Berman, Gopinath, Narasimhan, Yao
Studies verbal self-critique loops in which an agent reflects on prior attempts, generates feedback, and retries.
Madaan, Tandon, Gupta, Hallinan, Gao, Wiegreffe, Alon, et al.
Single-model iterative refinement via self-generated feedback. The minimal version of what multi-agent debate scales up across roles.
Wang, Bai, Liu, Chen, Cardie, Zhang, et al.
Layered multi-LLM collaboration where each layer's agents refine the previous layer's outputs. Open-source MoA reaches 65.1% on AlpacaEval 2.0, beating GPT-4 Omni.
Choi, Zhu, Li, et al.
Pinpoints when multi-agent debate actually beats majority vote: diversity-aware initialization plus calibrated confidence communication.
Swastik Raj
Practical walkthrough of orchestrating multi-agent debate in production — useful context for how debate systems are built and scaled in the wild.
Read on MediumPeer-reviewed methodology, explained for builders — free to read, no signup required.
A research-backed guide to choosing between self-consistency (CoT-SC) and multi-agent debate (MAD) — with benchmark numbers, cost analysis, failure modes, and a hybrid playbook for production.
Read guide →A practical playbook for adversarial AI deliberation — stress-test political messaging, product decisions, research claims, and legal arguments before they ship.
Read guide →Plain-English definitions for M-MAD, CoT-SC, Degeneration of Thought, sycophancy, and every term used across the research.
Browse terms →Start with a purpose-built workflow for decisions, expert review, debate, dialogue, team competition, pitches, qualitative research, audience simulation, idea selection, or a custom setup.
Two opposing agents pressure-test the assumptions and converge on a practical recommendation.
Output: Decision memo
A small panel debates your topic and distills the most useful insights.
Output: Key insights, no verdict
Two debaters build opposing cases while an arbiter scores the exchange.
Output: Arbiter's scored verdict
Two Worker-backed LLMs take turns in a private chat, each believing the other speaker is human.
Output: Dialogue transcript
Every member prepares privately, saved spokespersons debate in a fixed public sequence, and a neutral Arbiter scores the result.
Output: Five-dimension scorecard and verdict
Choose what is being decided, then have 2–4 Workers judge the pitch independently against that goal.
Output: Cited scorecard and verdict
A moderator follows an editable guide with known Worker personas and grounded findings.
Output: Themes and verified quotes
Generated participants react independently, then influence each other in networked rounds.
Output: Metrics and Observer report
Ideas are seeded, then argued head to head in a bracket until one survives with the best of the rest folded in.
Output: Champion spec and kill cards
Choose the protocol, participants, models, runtime, evidence, and notifications.
Output: Configurable
Attach source material to discussions. Truth-Seeking Debate can also prepare dedicated research for both sides and the neutral Arbiter.
Start with a focused pair or assemble a larger panel. Specialized formats add their own participant limits and roles.
Reuse configured participants, saved teams, visible role presets, shared rules, and hidden turn guidance across discussions.
Get a decision memo, scored verdict, qualitative report, audience metrics, dialogue transcript, or champion spec—depending on the format.
Copy a draft into hidden child runs, sweep sampling parameters, score each transcript, and stop on score, iteration, or cost limits.
Inject a human message into a running or paused discussion when the agents need a correction, constraint, or new piece of context.
Use cloud models through OpenRouter or local models through LM Studio, with configurable model fallbacks.
Bound work with turn, wall-clock runtime, and observed-cost limits so long-running discussions stop predictably.
Dedicated workers advance discussions, pitches, focus groups, simulations, evidence, and research outside long HTTP requests.
Review persisted turns, run logs, token and cost metadata, then export portable JSON records for further analysis.
Receive an in-app notification—and optional Resend email—when a run finishes or reaches its cost cap.
Multi Agent Debates includes dedicated workflows for pitch evaluation, moderated qualitative research, synthetic-audience simulation, and head-to-head idea selection.
Choose what is being decided, then have 2–4 Workers judge the pitch independently against that goal.
Cited scorecard and verdict
A moderator follows an editable guide with known Worker personas and grounded findings.
Themes and verified quotes
Generated participants react independently, then influence each other in networked rounds.
Metrics and Observer report
Ideas are seeded, then argued head to head in a bracket until one survives with the best of the rest folded in.
Champion spec and kill cards
Lab Experiments compare parameter choices across child runs, while Evidence Packs and optional Truth-Seeking internet research keep factual work connected to source material.
Start from a draft discussion, create hidden child runs, vary temperature and repetition, frequency, and presence penalties, then score each transcript against a validation prompt and expected outcome.
Upload an Evidence Pack for a discussion. In Truth-Seeking Debate, optional internet research prepares separate material for both debaters and neutral material for the Arbiter before the relevant phases run.
Compare runs without losing the transcript, parameters, evidence, or cost trail behind the result.
Multi Agent Debates keeps the transcript and operational trail, then adds the output that fits the job: a memo, verdict, report, scorecard, dialogue, metrics package, or tournament result.
Decision stress-tests produce a decision memo. Judged Debate returns an Arbiter score, and Team Battle ends with a five-dimension scorecard and verdict.
Focus Group produces grounded themes and verified quotes. Expert Panel distills key insights without forcing a winner.
Shark Tank returns a cited pitch scorecard and verdict. TribeMind provides descriptive metrics and a grounded Observer report.
Idea Tournament produces a champion spec and kill cards. Human Dialogue preserves the private-chat transcript, while Custom Discussion keeps its output configurable.
Multi Agent Debates can support message testing, qualitative research, pitch review, argument mapping, product decisions, audience simulation, and idea selection.
Use Team Battle for structured opposition, Focus Group for moderated reactions, or TribeMind to observe synthetic audience shifts across rounds.
Put a hypothesis in front of an Expert Panel or compare two competing interpretations in a Judged Debate with attached evidence.
Moderate known personas in a Focus Group, score a campaign pitch in Shark Tank, or simulate wider audience response with TribeMind.
Use a Decision Stress-Test or Judged Debate to compare competing cases, preserve the transcript, and keep supporting evidence with the run.
Challenge a roadmap choice with opposing agents, or send several concepts through an Idea Tournament for a champion spec and kill cards.
Review the transcript from an Expert Panel, 1:1 Human Dialogue, or Judged Debate and discuss how each participant handled the question.
Attach source material, run a Judged Debate or Team Battle, and collect the counterarguments your draft still needs to answer.
Put two AI personalities into a 1:1 Human Dialogue, build a custom discussion, or let several ideas compete in a tournament.
Quick answers to the questions teams ask before they run their first Multi Agent Debates session.
Multi-agent debate is an AI reasoning technique where multiple language models, or multiple instances of the same model with different roles, discuss a question across structured rounds. Research suggests it can improve some reasoning and evaluation tasks, though results depend on the models, format, and problem.
A single prompt gives you one model's first-pass answer. Multi Agent Debates gives you 10 purpose-built formats: Decision stress-test, Expert panel, Judged debate, 1:1 Human Dialogue, Team Battle, Shark Tank, Focus group, TribeMind, Idea Tournament, and Custom discussion. Each format preserves the run and produces an output suited to the job.
Multi Agent Debates supports models available through OpenRouter and local models served by LM Studio. You can select models per Worker and configure automatic fallback order in Settings.
The platform's debate workflows draw on published research from several institutions. The evidence is promising for some reasoning and evaluation tasks, but results remain model-, protocol-, and task-dependent. Key papers include:
We publish free, in-depth guides on multi-agent debate methodology—no signup required. Start here:
Use it to pressure-test decisions, gather specialist viewpoints, compare two arguments, observe a private AI-to-AI dialogue, stage Team Battles, evaluate pitches, run moderated Focus Groups, simulate audience response with TribeMind, select ideas in a tournament, or configure a discussion from scratch.
Choose 4–12 Workers with known personas, edit the discussion guide, and let a neutral moderator run the session. The resulting report is grounded in that simulated panel and includes themes and verified quotes rather than presenting the run as real population research.
Shark Tank returns a cited pitch scorecard and verdict. TribeMind runs 12–50 fictional participants through independent and social rounds, then produces descriptive metrics and a grounded Observer report. Idea Tournament compares 4–16 ideas head to head and produces a champion spec plus kill cards.
Lab Experiments copy a draft discussion into hidden child runs, vary temperature and repetition, frequency, and presence penalties, then evaluate each transcript against your validation prompt and expected outcome. They can stop on a score threshold, iteration limit, or total cost cap.
Multi Agent Debates focuses on purpose-built discussion formats, durable run monitoring, and format-specific reports rather than general-purpose agent orchestration. AutoGen, CrewAI, and LangGraph are broader frameworks for building custom agent applications and graphs.
Cost varies with the selected models, participant count, number of rounds, research calls, and report generation. Multi Agent Debates records token and cost metadata and supports turn, runtime, and observed-cost limits. Lab Experiments also support their own total cost cap.
Yes. All 10 discussion formats, Workers, Teams, Personas, Playbooks, Lab Experiments, live monitors, and exports are available through the web interface.
Multi Agent Debates does not train models on your transcripts. Prompts are sent to the model route you configure, so provider data handling depends on your OpenRouter model choice or LM Studio setup. The platform persists workspace records in its configured Supabase storage and provides archive, delete, and export controls.
Degeneration of Thought, formalized by Liang et al. (EMNLP 2024), is a failure mode where an LLM commits to an answer and struggles to produce genuinely novel reasoning during self-reflection. Separating roles across agents may help surface different arguments, but it does not eliminate the underlying risk.
Closed beta is rolling out gradually. Tell us what you would use it for, and we will send an invite when there is room.
Step 1 of 3
Your field