Exploring cognitive diversity in AI collaboration
Current AI assistants present a single perspective. This experiment explores whether multiple AI agents—each with distinct roles, mental models, and evaluation criteria—can produce better outcomes through structured discourse.
Participants run meetings where AI agents genuinely debate topics, surface disagreements, and reach decisions through visible reasoning.
Research Context
The observation
Single-agent AI tends toward agreement and accommodation. Real teams produce better outcomes through constructive disagreement between people with different expertise and priorities.
The hypothesis
AI agents with distinct mental models (how a CTO evaluates decisions vs. how a UX researcher does) will surface considerations that single-perspective AI would miss, and resist premature consensus.
The experiment
25+ agent archetypes across engineering, product, design, research, finance, and leadership. Each has codified evaluation criteria, blind spots, and communication patterns. Anti-sycophancy mechanisms prevent artificial agreement.
What Participants Do
Submit a topic for discussion
A product decision, technical architecture, research question, or problem to explore.
Select or accept a team
The system recommends relevant agents, or you can manually select participants.
Observe the discussion
Agents debate in real-time. You can interject to redirect or provide input.
Review artifacts
Decisions, action items, risks, and assumptions are captured as the meeting progresses.
Technical Approach
Agents are implemented across multiple LLM providers (Anthropic, OpenAI, Google) to introduce genuine model-level diversity. Each agent's prompt encodes a mental model with specific evaluation criteria, known blind spots, and non-negotiable positions.
Anti-sycophancy is enforced through 50+ communication patterns that prevent excessive agreement, gratitude, and accommodation. Agents are designed to maintain productive tension based on their role's actual priorities.
About this experiment
Teamwork is an early-stage research project, not a commercial product. The interface has rough edges. Agent behavior is being calibrated. Your participation and feedback contribute to understanding whether cognitive diversity in AI produces meaningfully different outcomes.