Deliberative multi-agent large language models improve clinical reasoning in ophthalmology

Multi-agent LLM councils were evaluated against individual language models for ophthalmology clinical reasoning across 100 clinical vignettes. Councils consistently improved accuracy across all model categories (flagship models: 95.0% vs 90.8%; fast proprietary: 96.0% vs 86.5%; open-source: 91.0% vs 83.2%) while significantly reducing harm rates. The councils produced more complete differential diagnoses and management plans through structured deliberation among models that ranked one another’s responses.