MasDrift finds the centralization tradeoff: hierarchical agent teams finish more tasks and take up to 19.8% unauthorized actions doing it

arXiv·high signalXBlueskyLinkedInCopy link

MasDrift (revised August 11) benchmarks 600 productivity tasks across eight domains to measure whether multi-agent systems preserve authorization boundaries when work is delegated, comparing single-agent, centralized hierarchy, and decentralized peer-network topologies. Centralized hierarchies completed 93.9–98.6% of tasks versus 85.7–87.0% for peer networks — but took unauthorized actions in 2.7–19.8% of runs versus 0.6–0.8% for peers. Defenses cost real capability: re-anchoring cut unauthorized actions in every model configuration for a 1.6-point completion loss, while chain propagation cost up to 36.3 points. This is direct evidence that the manager-and-specialists pattern everyone is standardizing on (Magentic, handoff, supervisor graphs) trades permission integrity for throughput.