Benchmark RadarAI BENCHMARK PROFILE
MasDrift
Evaluates multi-agent systems on 600 productivity tasks measuring task completion and unauthorized action rate across different coordination architectures.
- Released
- 2026-08-02
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Makes authorization preservation a measurable property of MAS design, exposing trade-offs between centralized and decentralized coordination for safe delegation.
Motivation
Multi-agent systems (MAS) decompose long-horizon tasks across supervisors and subagents, but delegated goals do not necessarily carry their original authorization boundaries.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.