When do multiple AIs or roles add value, and when do they merely repeat content?

Full explanation

Distinguish one model acting out experts, separate instances answering independently, and agents researching/executing different subtasks. These provide perspectives, answer comparisons and broader actual coverage, respectively, with different coordination and resource needs.

Du et al. had model instances answer independently and exchange views over rounds, improving some arithmetic and knowledge-QA results. ICML paper Cemri et al.'s multiagent trace analysis found failures in work design, inconsistent information, verification and termination. Failure analysis Consensus is a process outcome, not proof of correctness.

Anthropic's research-system engineering report likewise finds more benefit in parallel exploration than tightly dependent work requiring extensive shared context. This is its product experience. Official report

An arrangement could be:

Assign three agents to sources, calculations and reader comprehension. Have them form findings separately before the lead synthesizes them. Each issue needs location, evidence and recommendation; merge duplicates and preserve evidenced disagreements. The lead verifies key facts and delivers one revised draft.

For independent problem solving, form answers before exchange to reduce following the first. For execution, define artifacts/interfaces to avoid simultaneous overwrites. Assess added evidence/corrections, total time and cost rather than retaining repetitive statements for a sense of teamwork.

References

1. Du et al. (2024), Improving Factuality and Reasoning in Language Models through Multiagent Debate, ICML

2. Cemri et al. (2025), Why Do Multi-Agent LLM Systems Fail?, initial trace analysis and failure taxonomy

3. Anthropic (2025), How we built our multi-agent research system, Benefits of a multi-agent system