Researchers have proposed an always-on intelligence system called Jarvis that acts as a mediator between humans and agents working in the background. This system is designed to allocate human attention across multiple agents and provide immediate access to user-initiated questions about ongoing work. To evaluate this concept, the authors created JarvisBench, a benchmark consisting of 45 task instances from various domains, which tests both directions of coordination: whether an intermediary can accurately answer user queries and recognize when an agent requires user judgment. The system is designed to integrate with arbitrary agent runtimes without modifying their execution loops. --- Why it matters: This matters because it addresses the issue of human attention scarcity in long-horizon agents, allowing for more efficient and effective collaboration between humans and AI systems. Source: https://arxiv.org/abs/2608.14870