A newly posted research paper on Hugging Face argues that proactive LLM agents — systems that act before users ask — can convert idle compute into useful assistance, but also risk misreading user context, imposing review burdens, or eroding trust. The authors propose a foundation for designing, building, and evaluating such agents around three joint principles they call the 3T: Task Capability, Temporal Allocation, and Trust.
Task Capability refers to anticipating relevant needs and correctly performing useful work. Temporal Allocation concerns distributing compute according to resource availability and when results are needed. Trust focuses on sustaining users' confidence and appropriate reliance on the agent. The paper connects these objectives to a design space organized around five dimensions: task scope, anticipation horizon, activation trigger, processing timing, and intervention depth.
The authors also specify the situation and system modeling needed to support design choices, including user and environment representations, backbone LLMs, and agent harnesses. To evaluate proactive behavior, they introduce PROACTIVITY-GYM, a simulation-based testbed featuring multi-day scenarios, stateful environments, and persona-conditioned simulated users that can assess the consequences of proactive assistance across interactions.
Evaluations across 23 model-harness configurations reportedly uncover substantial performance gaps across the 3T principles and indicate that LLM judges often conflate task capability with trust. A human study with 30 participants highlights the importance of joint 3T optimization: participants showed sharp trust declines after intervention misalignment despite correct outcomes, and preferred sleep-time assistance — even when imperfect — to preserve ongoing focus.
The paper concludes that proactive agents should be designed and evaluated through the joint consideration of useful work, compute allocation, and evolving user trust.