'Multi-Agent Systems Underperform a Single Agent With Full Information' — 119-Point HN Comment Leads the Pushback on Anthropic's Research
The top-voted comment (119 points) on Anthropic's multi-agent post argues the whole framing buries the finding that a single agent with complete information beats the distributed setup, making the coordination research a solution in search of a problem. A second recurring thread attacks the conformity result from the other direction: LLM output clusters around the dataset average, so a swarm of identically-trained agents has structurally less variance than a human team and cannot produce the diversity the paper prescribes. Several commenters dismiss it as 'a promo article, aka an ad' and 'a desperate attempt by Anthropic to meta-explain away deficiencies,' while the substantive agreement is narrow — that hierarchical, domain-specialist subagents with constrained action spaces beat flat peer arrangements. For anyone running an agent fleet, the thread is a useful corrective: the default should be one well-informed agent, with fan-out justified per task, not assumed.
Source
↳ Follow the thread