A study on multi-agent LLM collaboration found that upstream messages can both help and hurt downstream agents. Messages can improve accuracy, but also lead to incorrect answers 32% of the time. The study suggests that communication should be selective based on upstream reliability and evidence.