Stateful assistants stay below 70% completion

State-of-the-art proprietary assistants failed to reach 70% task completion on scenarios requiring persistent state and configuration awareness in the PAUSE benchmark. Its tasks span user-owned resources, permissions, multiple services, and multi-turn interactions, exposing failures that static tool calls conceal. As enterprises assemble agent gateways around hundreds of tools and non-human identities, builders should make state and authorization checks explicit acceptance gates before assistants can change consequential settings.