Multi-Agent Brainstorming: How Simulating a Debate in One Chat Beats a Human Team at Ideas

All steps in one prompt, one request:

ROUND 1:
  Agent A → proposes an idea (a wide leap into a new domain)
  Agent B → compares it with A's idea, picks the better one or creates a hybrid
  → Best idea of the round locked in

ROUND 2:
  Agent A → proposes a NEW idea from a different domain (doesn't build on the previous one)
  Agent B → compares it with the best from round 1, picks the winner
  → Best idea of the round locked in

ROUND 3:
  Agent A → one last leap, into an even more unexpected domain
  Agent B → final pick
  → RESULT: the strongest idea by novelty and practicality

The whole process is one prompt, one response from the model.

Task: Come up with a monetization format for a personal finance Substack newsletter with 15,000 subscribers. The standard formats (sponsorships, courses, consulting) are already oversaturated.

Prompt:

You are simulating a discussion between two agents, Agent A and Agent B, 
on this task: {come up with an unconventional way to monetize 
a personal finance Substack newsletter, 15,000 subscribers, 
the niche is flooded with courses and sponsorships}.

Discussion rules:
— Each round, Agent A proposes an idea from a completely different industry 
  or an unexpected context (does NOT build on the previous one; leaps somewhere new)
— Agent B compares the new idea with the best previous one, picks the stronger 
  or creates a hybrid, and explains the choice in one sentence
— After each round, record the "Best idea of the round"

Run 3 rounds. At the end, give the final idea with a short justification: 
why it is both novel AND practically doable.

Result:

The model will show three rounds of discussion with clear jumps between conceptual domains, for example from subscription-service mechanics to gamification, then to a DAO or crowdfunding model. Each time, Agent B will briefly explain why it picks one idea over another. At the end you get a concrete idea with a justification for why it's both non-obvious and realistic.

The LLM's weakness in "free" mode: the model outputs the first tokens that fit the request and then keeps "drifting" in the same direction. Ask for "10 ideas" and you get 10 variations of one idea, just in different words. That's not a bug; it's how text generation by next-token probability works.

The LLM's strength: it can simulate roles with different mindsets and make big semantic leaps without any cognitive discomfort. A person feels awkward suggesting an idea that seems "too weird"; an LLM has no such barrier. Novelty grows out of exactly those leaps.

The iterative refinement structure takes advantage of this: a forced move into a new domain every round ensures wide coverage of the semantic space, while the "compare and pick" mechanic keeps practical value in check, so ideas don't float off into abstraction.

Prompt control levers: - Number of rounds → 2–3 for quick tasks, 5+ for strategic ones - How you phrase the leap → the more specific the forced instruction ("from a completely different industry"), the farther the leap and the higher the novelty - Agent B's selection criterion → you can add a second criterion, for example "doable within 30 days" - Number of agents → the study shows: 3 beats 6 (coordination losses wipe out the gain)

You are simulating a discussion between two agents, Agent A and Agent B, 
on this task: {describe the task}.

Discussion rules:
— Each round, Agent A proposes an idea from a completely different 
  domain or context (does NOT build on the previous one; leaps somewhere new)
— Agent B compares the new idea with the best previous one, 
  picks the stronger or creates a hybrid, 
  and explains the choice in one sentence
— After each round, record the "Best idea of the round"

Run {number_of_rounds} rounds. 
At the end, give the final idea with a justification: 
why it is both {criterion_1} AND {criterion_2}.

What to fill in: - {task}: a concrete statement: what to create, what to name, what to solve, what to come up with - {number_of_rounds}: 3 for most tasks - {criterion_1}: novel / non-obvious / unconventional - {criterion_2}: realistic / doable / commercially viable

🚀 Quick start: paste this into your chat:

Here is a multi-agent brainstorming template. 
Adapt it to my task: [your task]. 
Ask me questions to clarify the details.

[paste the template above]

The LLM will ask what exactly you need to come up with and which criteria matter more, because without that it can't tune Agent B's selection criterion properly. It will take the pattern from the template and adapt it to your task.

Intuition says: give the agents different roles ("optimist", "skeptic", "engineer") and the ideas will get more diverse. The study disproves this.

Assigning personas doesn't raise overall creativity. What it does is shift the balance: agents with personas produce less novel but more practical ideas. Agents without personas do the opposite: weirder, more unexpected ideas.

The practical takeaway: - Need maximum novelty → drop personas entirely - Need a novelty + practicality balance → add personas with specific constraints ("keeps the budget in mind", "focuses on execution")

⚠️ The ratings were done by people: Human experts rated the ideas on novelty and usefulness. If your criterion is subjective or niche-specific, add it to the prompt explicitly; otherwise "novelty" will be judged in the general sense.

⚠️ For powerful models (o3, Claude Opus), structure matters less: The study found that reasoning models with deep internal logic gain almost nothing from external discussion structures. If you're working with a top-tier model, you'll get good results even without the multi-agent template. The structure helps standard models (GPT-4o and the like) the most.

⚠️ Don't overcomplicate routine tasks: If the task is simple ("write a photo caption" or "write a birthday message"), the multi-agent structure is overkill. The method pays off where you need genuine non-obviousness, not just useful text.

⚠️ Usefulness can drop as novelty rises: The study recorded a trade-off: structures that sharply boost novelty slightly reduce how practical the ideas are. Keep this in check through Agent B's selection criterion.

A team from the University of Cambridge and Microsoft Research took six real-world problems, from fighting plastic waste to tackling employee turnover at companies, and collected 4,541 ideas from multi-agent LLM teams and 341 ideas from 114 real human teams. Five independent experts rated the ideas blind, without knowing whether a person or a model wrote them.

An interesting design choice: the tasks deliberately included hypothetical scenarios (like "a pandemic in which people sing in the shower") to rule out the LLM simply "remembering" the right answers from its training data. Even on those tasks, the LLM teams won.

A surprising result: large teams (6 agents) didn't outperform small ones (3 agents), for humans or for LLMs. Too many participants create coordination losses that eat up the entire gain from diversity. It's the classic "process loss" from social psychology, and AI reproduced it exactly the way people do.

Another fascinating finding: to understand why LLM teams win, the researchers represented each conversation as a trajectory in semantic space and literally measured the "length", "width" and "smoothness" of the path the conversation takes. It turned out that highly creative LLM conversations made sharp, wide leaps between conceptual domains (for example, from "economic incentives" through "blockchain" to "biotech") in the fewest possible steps. Low-creativity conversations circled around a single idea.

🔧 Add an explicit instruction on leap width

By default, an LLM may "leap" into neighboring domains rather than distant ones. Push it further:

Each round, Agent A MUST take an idea from a domain 
that at first glance has nothing to do with the task: 
a different industry, a different culture, a different scale (from an atom to a city), 
a different era. The farther the leap, the better.

Effect: the conversation literally "sweeps" wider and novelty goes up.

🔧 Control the novelty/practicality balance through Agent B's criterion

The standard template gives novelty and usefulness roughly equal weight. You can shift that:

If you need maximum novelty:

Agent B picks the idea that is more UNEXPECTED and unconventional, 
even if it's harder to execute

If you need maximum feasibility:

Agent B picks the idea that can be launched 
within 2 weeks with zero budget, even if it's less original

🔧 Extrapolation: a multi-agent review before a big decision

The same mechanics work not just for generating ideas but also for analyzing decisions:

You are simulating a discussion between two analysts, Analyst A and Analyst B, 
on this question: {decision or situation}.

— Each round, Analyst A offers an unexpected angle 
  (from another discipline, another scale, another time horizon)
— Analyst B compares it with the previous angle and picks 
  the more valuable one or creates a synthesis

Run 3 rounds. Finale: the least obvious but most important insight about the situation.

Works for: evaluating a business idea, thinking through a career decision, finding risks in a plan.

Multi-agent AI systems outperform human teams in creativity (2025)

Authors: Tiancheng Hu, Yixuan Jiang, Haotian Li, José Hernández-Orallo, Xing Xie, Nigel Collier, David Stillwell, Luning Sun

Organizations: University of Cambridge (Department of Theoretical and Applied Linguistics; The Psychometrics Centre, Cambridge Judge Business School; Leverhulme Centre for the Future of Intelligence), Zhejiang University, Microsoft Research Asia, Universitat Politècnica de València (VRAIN)

Author contact: [email protected]