Anthropic's 'Worse Together': single coordinator agent beats per-user agent teams in all 4 tests
ombre-ai · reddit · 2026-11-12
Anthropic's Alignment Science blog (Anthropic Fellows research) compared per-user agent teams sharing one limited resource against a single coordinator across 4 environments, 5 frontier models, and 77 scenarios. The coordinator won every time; without a communication channel, teams collapsed entirely in half the environments, and observed failures included stalling, agents overriding users, and unsupported claims. Caveats: simulated environments only, and a coordinator is itself a bottleneck and single point of failure.
More from coding & agent
- Qwen's terminal coding agent qwen-code nears 30k GitHub stars — QwenLM · 2026-11-12
- Dev uses RF-DETR to digitize tabletop wargames, shares IR frame pitfalls — TheZachMueller · 2026-11-12
- Developer cuts agent token costs to $0.24 per session, comfortably runs Opus 5.5 — pramodk73 · 2026-11-12
- Alibaba open-sources Open Code Review, its internal AI code reviewer with 50K GitHub stars — techNmak · 2026-11-12
- Claude Code built a toy-town sim of Fürth where every train runs on real timetable and sensor data — Greedy-Giraffe-4269 · 2026-11-12
- Community-built desktop app runs Grok Build, Codex, Claude Code agents without an IDE — PawelHuryn · 2026-11-12