Designing for Productive Disagreement: How to Preserve Exploration Before Synthesis
An AI brainstorming workflow should be designed around the point at which exploration is most likely to disappear. This chapter turns the evidence about fixation, personas, authority, and communication topology into an operating pattern: let agents generate independently, vary the cues or models that shape their search, introduce critique only after alternatives exist, and verify both breadth and usefulness before synthesizing. The goal is not to keep a team disagreeing forever. It is to make later agreement an informed choice rather than an automatic consequence of shared context.
Build independence into the sequence
The first design decision is whether the task actually needs broad exploration. The evidence concerns open-ended fitness-product ideas and scientific research proposals, where distant alternatives can matter. A problem with a strong ground truth may call for convergence, and forcing extra semantic variation in such a task can create noise or hallucination. Before adding agents or prompts, define what a valuable alternative would change: a customer need, a product form, a research mechanism, an assumption, or a method. This keeps diversity connected to the subject instead of turning it into a target detached from usefulness.
For an exploratory task, begin with independent first passes. Do not let every agent see the first answer before producing a starting proposal. Shared visibility makes an early idea salient and gives the group a common frame before the members have contributed their own regions of the search space. The human–LLM experiments show why this matters: first ideas are a relatively clean indicator of starting-point variation, while later ideas are already affected by a session’s history. A blind first pass protects the measurement and the search at the same time.
Independence does not mean asking identical agents to work in silence. The point is to create meaningful differences among the initial searches. Ordinary heterogeneous personas are one evidence-backed option. In the fitness study, varied everyday backgrounds and professions sent sessions toward more separated semantic regions than default prompts or curated creative-entrepreneur archetypes. A persona should be treated as a sampling cue that changes what a session notices and values, not as a fictional claim that the model has acquired a person’s real experience.
Where the operational cost is acceptable, genuinely different models are another option. The multi-agent experiments found that mixing DeepSeek-V3, GPT-4o, and Claude-Sonnet-4 increased diversity in authority-weighted structures, where same-model agents tended to converge. Different training data, architectures, alignment objectives, and reasoning styles can supply different priors. But model heterogeneity is not automatically superior. It can create inconsistent assumptions, require more synthesis, and add cost. It should be evaluated as a source of potentially independent information, not as a badge of sophistication.
The first pass should also preserve the identity of the alternatives. Keep each proposal associated with the cue or model that produced it until comparison is complete. If everything is immediately merged into one shared conversation, the system loses the ability to tell whether two ideas were independently discovered or merely elaborated after one agent suggested them. This is especially important when the final output is polished by an editor: synthesis should combine visible alternatives, not conceal the paths that produced them.
The next stage is controlled sharing. Once independent proposals exist, agents can inspect one another’s work to identify contradictions, missing assumptions, and promising combinations. The order matters. Sharing is valuable because it enables critique and recombination; sharing too early is dangerous because it turns the first contribution into an agenda.