MACI: Scheduling Confrontation in Multi-Agent Debates

STEP 1 (CL=0.9): Exploration
  Agents A and B get the task → give independent opinions
  Instruction: "Criticize your opponent aggressively, look for weaknesses"
  Moderator: lets through any argument above a minimum threshold

STEP 2 (CL=0.7): Integration
  Agents get each other's arguments → refine their positions
  Instruction: "Criticize in a targeted way, acknowledge strong points"
  Moderator: raises the evidence quality threshold

STEP 3 (CL=0.5): Consolidation
  Agents → synthesize an agreed conclusion
  Instruction: "Build on common ground, focus on the evidence"
  Moderator: checks for a plateau (disagreement <5%, information gain <2%)

STOP: if disagreement and information gain don't change for 2 rounds in a row

Every step runs as a sequence of prompts. In each round the moderator (that can be you, or an LLM with its own prompt) checks: are new arguments showing up? Are the conclusions converging? If nothing's changing — stop.

Task: Evaluate a startup idea — a delivery service for ready-made healthy meal plans for office workers in Chicago (think "Factor" for corporations).

Round 1 prompt (CL=0.9 — exploration):

I have a startup idea: delivering ready-made healthy lunches
to offices. Target audience: tech and consulting firms, price
$14-18 per lunch, weekly subscription format.

I need TWO independent opinions:

AGENT A — the optimist-visionary (Mark Cuban style):
Find every opportunity, describe the upside potential, ignore the risks.

AGENT B — the skeptical investor (Kevin O'Leary style):
Find every risk and problem, challenge every assumption.

CRITICALLY IMPORTANT: criticize each other aggressively,
don't look for compromises. Each of you, give 3-4 arguments
with specific numbers/examples.

Result: The model returns two contrasting takes. Agent A talks up the growing healthy-eating market (30% CAGR), willingness to pay a premium, network effects. Agent B points to fierce competition (Factor, ezCater), tricky logistics, thin margins (10-15%), high CAC.

Round 2 prompt (CL=0.7 — integration):

Here are Agent A's arguments: [paste]
Here are Agent B's arguments: [paste]

Now RESPOND TO EACH OTHER:
- Agent A: acknowledge 1-2 strong objections from B, but explain
  how they could be worked around
- Agent B: acknowledge 1-2 real opportunities from A, but name
  the specific conditions under which they'd pan out

Bring EVIDENCE (case studies, data, calculations),
not generalities.

Result: The positions start converging. Agent A concedes the unit economics problem but proposes a hypothesis: focus on large offices (100+ people) with a minimum order → lower delivery cost. Agent B agrees there's margin in the premium segment, but notes you'd need a pilot with 2-3 corporate clients to test retention.

Round 3 prompt (CL=0.5 — synthesis):

Agents A and B — synthesize an AGREED conclusion:

1. TOP 3 critical assumptions that a pilot needs to test
2. One concrete scenario in which the idea WORKS
3. One concrete scenario in which the idea FAILS

Build ONLY on the arguments that you both found
convincing. Ignore weak hypotheses.

Result: A consolidated conclusion with clear success conditions (retention >40%, AOV >$150/month, CAC <$60) and testable hypotheses (willingness to pay for a subscription, order frequency, operating margin). If the agents keep arguing, add round 4 with CL=0.3 ("focus only on common ground").

Single LLMs suffer from first-impression lock-in. A model generates a probabilistic output based on patterns from its training data — if the first hypothesis sounds plausible, the model will rationalize it instead of checking it critically. This shows up as confirmation bias at the token-generation level. One model can't simultaneously explore alternatives and defend a position — those are contradictory modes.

LLMs are good at role-play and argumentation. When the prompt explicitly assigns a role ("skeptic," "optimist"), the model activates the matching patterns from its training data. It's not "understanding" the role, it's statistical activation of contexts where such roles appeared (debates, expert panels, scientific discussions). Multiple agents with different roles widen coverage of the hypothesis space — each agent explores its own territory instead of everyone converging on the same local maximum.

Scheduling confrontation avoids two traps: (1) premature consensus — agents agree on the first plausible idea without exploring alternatives; (2) endless arguing — agents keep generating new objections after all the strong arguments have already been made. High confrontation at the start (CL=0.9) forces models to actively hunt for counterexamples and problems. Low confrontation at the finish (CL=0.3-0.5) switches the mode to integrating evidence instead of generating new hypotheses.

Raising the quality bar works as a noise filter. Early on, even weak hypotheses are useful — they widen the search. Later, weak arguments only muddy the conclusion — the evidence quality threshold rises as things converge. It's analogous to simulated annealing in optimization: start with a broad search (high "temperature"), finish with local refinement (low "temperature").

Simplified version (3 rounds)

# ROUND 1: Exploration (CL=0.9)

{task description}

I need TWO independent opinions:

AGENT A — the {role_A, e.g.: optimist/visionary}:
{instruction for A}

AGENT B — the {role_B, e.g.: skeptic/critic}:
{instruction for B}

CRITICALLY IMPORTANT:
- Criticize each other aggressively
- Don't look for compromises
- Each of you, give 3-5 specific arguments with evidence

---

# ROUND 2: Integration (CL=0.7)

Here are Agent A's arguments: {paste A's output from round 1}
Here are Agent B's arguments: {paste B's output from round 1}

Now RESPOND TO EACH OTHER:
- Agent A: acknowledge 1-2 strong objections from B, explain how to work around them
- Agent B: acknowledge 1-2 real opportunities from A, name the conditions
  under which they'd pan out

Bring EVIDENCE (data, case studies, calculations), not generalities.

---

# ROUND 3: Synthesis (CL=0.5)

Agents A and B — synthesize an AGREED conclusion:

1. {key question 1, e.g.: critical assumptions}
2. {key question 2, e.g.: success scenario}
3. {key question 3, e.g.: failure scenario}

Build ONLY on the arguments that you both found
convincing. Ignore weak hypotheses.

IF YOUR OPINIONS STILL DIVERGE → move on to Round 4 with CL=0.3

Notes on the placeholders:

• {task description} — your problem/question
• {role_A}, {role_B} — specific roles (optimist/pessimist, engineer/business person, conservative/innovator)
• {instruction for A/B} — what each agent should look for
• {key question N} — what the final synthesis must contain (success criteria, risks, an action plan, etc.)

Manual stopping rule: If in round 3 the agents produce nearly identical conclusions (<10% disagreement on the key points), you're done. If they keep arguing, add round 4 with an even softer instruction: "focus ONLY on common ground, ignore the disagreements."

⚠️ Takes multiple requests: At least 3 rounds = 3 separate prompts. Complex tasks may need 4-5 rounds. If you're limited to a single request or need a quick answer, the method is overkill.

⚠️ Quality depends on the underlying models: With a weak model (GPT-3.5, early versions of Claude), the agents may produce shallow arguments even with the right confrontation schedule. The method works better with modern models (GPT-4, Claude 3, Gemini Pro).

⚠️ Overkill for simple questions: If the task has an obvious answer or only a few options (choosing between 2-3 alternatives), a debate won't add anything. A simple "give me two opinions" will be faster. MACI is useful for multidimensional problems with no single right answer.

⚠️ Subjective quality criteria: The study used an automated moderator with metrics (Jensen-Shannon divergence, mutual information). Doing it by hand, you'll have to judge for yourself when the disagreement is small enough and when new arguments have stopped coming. That takes attention.

The researchers took 1,500 medical cases (30 diseases) and tested three pairs of LLM agents: GPT-4o + Claude 3 Opus, GPT-4o + Gemini Pro, Claude + Gemini. Each pair got a patient's symptoms and had to reach a diagnosis through 2-4 rounds of debate.

They compared against three baselines: the best single model, majority voting (20 independent answers), and debates with fixed confrontation (the Liang et al. method). For each case they checked by hand: (1) Top-1 and Top-3 diagnostic accuracy, (2) calibration (how well the model's confidence matches its accuracy), (3) token count.

Why the results came out the way they did: The medical cases were deliberately chosen to be underspecified — only current symptoms, no lab results or history. This mimics the real situation where a doctor has to order initial tests without the full picture. Under these conditions majority voting falls flat (73.9% accuracy), because all the models lock onto the same first hypothesis (say, "flu" for a high fever). MACI forces the agents to explore alternatives through confrontation (round 1), then integrate the evidence (rounds 2-3), and finally synthesize while accounting for uncertainty (rounds 3-4).

The surprise: The biggest accuracy gain came not from the number of agents but from the behavior schedule. When scheduling was switched off (agents stayed either high-confrontation or low-confrontation the whole time), accuracy dropped by 3.9-6.0 percentage points. This shows that timing the confrontation matters more than simply "getting more opinions."

They also tested on 619 news articles (political scandals with Democrat/Republican partisan ratings). Here the agents started from opposite partisan framings and had to arrive at a centrist assessment. MACI narrowed the partisan gap by 68% vs. the baseline methods, using the same hyperparameters as for medicine — confirming that the scheduling principle transfers across domains.

💡 Cutting it down to 2 rounds for quick tasks

If your task is simple (choosing between 3-4 options, no deep analysis needed), you can cut it down to two rounds:

Round 1 (CL=0.9): Two contrasting opinions, no compromises Round 2 (CL=0.5): Synthesis based on the strong arguments

Example prompt:

# ROUND 1: Exploration
Give two contrasting opinions on the question {question}:
- Opinion A (supporter): every argument FOR
- Opinion B (opponent): every argument AGAINST
Criticize aggressively, give specific examples.

# ROUND 2: Synthesis
Agents A and B — synthesize a consensus conclusion:
which arguments do you both find convincing?

This saves one prompt but keeps the key principle: high confrontation → low confrontation.

🔧 Technique: Swap faceless agents for personas

What changes: Instead of "Agent A" and "Agent B," use recognizable personas with a distinctive way of thinking.

The effect: The model activates more specific behavior patterns. Instead of a generic "optimist/pessimist," you get the specific framing characteristic of that person.

Example:

I need two opinions on the idea {description}:

MARK CUBAN — an aggressive optimist, focused on scale and speed:
Find the upside potential, ignore the risks, think about how to win the market.

KEVIN O'LEARY — a conservative investor, focused on unit economics:
Find every hole in the business model, run the numbers, test the assumptions.

ROUND 1 (CL=0.9): Criticize each other aggressively.
ROUND 2 (CL=0.5): Synthesize based on the strong arguments from both.

Other useful persona pairs:

• Engineer vs. Business person: Elon Musk (innovation) vs. Warren Buffett (conservatism)
• Marketer vs. Product person: Philip Kotler (market) vs. Steve Jobs (product)
• Policy: a reform-minded activist vs. a cautious career regulator — for analyzing policy decisions

Personas make the role-play sharper because the model draws not on abstract "optimism" but on the specific speech patterns and reasoning of a real person.

💡 Adapting it for personal decisions

Context: Instead of business analysis, apply it to personal dilemmas (moving, changing jobs, a big purchase).

Example prompt:

I need to decide: {description of the dilemma, e.g.: move from Chicago
to Austin or stay}

Give me two opinions:
- Agent A (pro-move): every argument FOR moving
- Agent B (pro-staying): every argument AGAINST moving

ROUND 1 (CL=0.9): Criticize aggressively, look for holes in your
opponent's logic. Give specific examples and numbers (cost of living,
salaries, quality of life).

ROUND 2 (CL=0.7): Agent A — acknowledge 2 strong risks raised by B.
Agent B — acknowledge 2 real opportunities raised by A.

ROUND 3 (CL=0.5): Synthesize:
1. Under what conditions does moving make sense?
2. Under what conditions is it better to stay?
3. What information do I need to gather to decide?

Why it works for personal decisions: Big personal decisions often suffer from emotional reasoning — we look for confirmation of what we already want to do. Two contrasting agents force you to spell out both sides with concrete arguments, not just "I want to / I don't want to."

Original paper: Multi-Agent Collaborative Intelligence: Dual-Dial Control for Reliable LLM Reasoning (2025)

Authors: Edward Y. Chang, Ethan Y. Chang
Affiliation: Stanford University, University of Illinois Urbana-Champaign (UIUC)

Related work cited in the study:

• CRIT (Socratic evaluation protocol) — a method for evaluating arguments through cross-questioning
• Liang et al. (2024) — fixed confrontation in multi-agent debates
• Wang et al. (2023) — Self-Consistency for LLMs
• Mixture-of-Agents (Wang et al., 2024) — ensembling through layers of agents