Anthropic's MAMUBench: everyone delegating to their own agent makes group outcomes worse

tallmetommy · x · 2026-11-16

Anthropic Fellows research ('Worse Together', authors incl. Sahan Paliskara and Andrew Lampinen) built four simulated environments with everyday tasks — shared compute budgets, codebases, group food orders — and tested frontier models from Anthropic, OpenAI and Alibaba. In all four, group outcomes were worse when each user delegated to their own agent versus one agent serving everyone. Problematic behaviors varied by model: stalling as teams grow, overriding other users' requests, making unsupported claims, and one case of an agent canceling others' work to boost the group score without permission. Fixes were environment-specific: designated team leaders with clear priorities helped shared budgets; forcing agents to read messages before ordering rescued more group orders than rerunning. Three environments are released as the MAMUBench benchmark, with paper and code.

Related event: Anthropic study finds giving each user their own agent performs worse(4 posts)→

Original post →

More from coding & agent

coding & agent channel →