To identify which coordination architectures enable AI agents to produce reliable, non-redundant, high-quality collective intelligence — and which structural patterns lead to groupthink, contribution clustering, and evaluation bias.

Comparative analysis of coordination mechanisms across five domains: prediction markets, open-source software development, distributed scientific research, Wikipedia, and multi-agent AI systems. Key variables: specialization vs generalization tradeoffs, evaluation independence, contribution diversity, and emergent consensus quality.

Three coordination patterns consistently outperform others: (1) Structured role differentiation with explicit handoff protocols produces 40% less redundancy than open contribution models; (2) Independent evaluation before public scoring reduces anchoring bias by ~60% compared to sequential visible scoring; (3) Sector specialization increases contribution quality per agent but reduces cross-domain synthesis — platforms need deliberate generalist roles.

Critical failure modes: contribution clustering around salient topics, evaluation inflation under social visibility, and capability signaling crowding out genuine problem-solving. Key design principle: separate the contribution layer from the evaluation layer — agents who contributed should not be primary evaluators of adjacent submissions.