Anthropic simulations highlight coordination risks in agent teams
The November 11 research compares agents sharing resources; its findings are not universal real-world team rankings.
Anthropic, Fellows Program and independent researchers published a multi-user, multi-agent study on November 11, 2026, comparing formations in four simulated environments. They report poorer group outcomes for per-user agent teams than a single coordinator in tested settings.
This is vendor-involved simulated research, not a universal verdict on multi-agent systems. Evaluators should examine authorization, message handling and shared-resource actions alongside task outcomes; completion scores alone cannot establish safety.
The experiments were not independently reproduced here. Chinese and Cantonese scenarios were not verified. This research release makes no new product-pricing or Hong Kong, Macau or Taiwan account-availability commitment.
References:
https://alignment.anthropic.com/2026/worse-together/
This is vendor-involved simulated research, not a universal verdict on multi-agent systems. Evaluators should examine authorization, message handling and shared-resource actions alongside task outcomes; completion scores alone cannot establish safety.
The experiments were not independently reproduced here. Chinese and Cantonese scenarios were not verified. This research release makes no new product-pricing or Hong Kong, Macau or Taiwan account-availability commitment.
References:
https://alignment.anthropic.com/2026/worse-together/
Usage conditions
- Hong Kong
- This is simulated research rather than a regional product launch. Chinese and Cantonese scenarios and local product availability were not established.
- Macau
- This is simulated research rather than a regional product launch. Chinese and Cantonese scenarios and local product availability were not established.
- Taiwan
- This is simulated research rather than a regional product launch. Chinese and Cantonese scenarios and local product availability were not established.
Official source: Worse Together: How Performance Breaks Down in Multi-User Multi-Agent Teams