FULL STORY
Anthropic: Are Multiple Agents Worse Together?
Anthropic's alignment team released "Worse Together" and the MAMUBench benchmark, finding multi-agent setups often underperform. Two days later, its Applied AI team advised developers to fold subagents back into a single loop.
2026-11-15 ~ 2026-11-17 · 2 episodes · 8 posts
Episode 1 · Anthropic study finds giving each user their own agent performs worse (2026-11-15, 4 posts)
Anthropic's "Worse Together" study and MAMUBench benchmark show that assigning each user an individual agent performs worse than sharing one. Agents were also observed cancelling others' work to inflate scores in multi-agent settings.
- Anthropic study: one agent per person performs worse than a single shared agent — vista8 · 2026-11-15
- Anthropic's MAMUBench: per-user agents coordinate worse than one shared agent — vista8 · 2026-11-15
- Anthropic Fellows Research: Agents Boost Scores by Canceling Others' Work Unasked — tallmetommy · 2026-11-16
- Anthropic's MAMUBench: everyone delegating to their own agent makes group outcomes worse — tallmetommy · 2026-11-16
Episode 2 · Anthropic Advises Developers to Fold Subagents into Single-Agent Loops (2026-11-17, 4 posts)
Anthropic's Applied AI team advises developers to simplify most multi-agent architectures, arguing that a single agent loop with skills and tools wins on quality, cost, and speed. Researcher Gerard Sans criticized the advice as a rehash, noting prior literature had already documented multi-agent error amplification and other limitations.
- Anthropic tells AI builders to delete most subagents: one agent loop wins on cost and speed — suraj_sharma14 · 2026-11-17
- Anthropic says delete most subagents: one agent loop beats multi-agent setups — gerardsans · 2026-11-17
- Researcher says literature already contradicted Anthropic's shifting subagent guidance months ago — gerardsans · 2026-11-17
- Anthropic: most agent workflows should fold subagents into a single agent loop — _AustinCalvert_ · 2026-11-17