Foundations of Proactive Agents: Principles, Technical Layers, and Proactivity-Gym

Foundations of Proactive Agents. Principles, Technical Layers, and Proactivity-Gym. By Jio Oh and four others, from KAIST and the University of Minnesota. Published on the twenty-ninth of September. No code repository is linked to it.

Proactive LLM agents can turn idle compute into useful support before a user even asks. But doing the work correctly is not the whole job. Even correct work can misread the user's context, impose the cost of reviewing it, or undermine the user's trust. Judging a proactive agent only by whether its output is right misses those consequences, and they play out across interactions rather than inside a single exchange.

The work proposes foundations for designing, realizing and evaluating such agents, built around three joint principles, referred to as three T. Task Capability means anticipating relevant needs and correctly performing useful work. Temporal Allocation means allocating compute according to the resources available and to when results are needed. Trust means sustaining the user's confidence, along with an appropriate reliance on the agent. The word joint matters: the three are meant to be pursued together, not traded off one at a time.

These objectives are connected to a design space with five dimensions: task scope, anticipation horizon, activation trigger, processing timing, and intervention depth. The authors also specify the situation and system modelling needed to support choices in that space, including representations of the user and the environment, the backbone LLMs, and the agent harnesses.

For evaluation there is Proactivity-Gym, a simulation-based testbed. It includes multi-day scenarios, stateful environments, and simulated users conditioned on personas, so that the consequences of proactive assistance can be evaluated across interactions, as they unfold, rather than one response at a time.

Evaluations across twenty-three model and harness configurations uncover substantial performance gaps across the three principles. They also reveal that LLM judges often conflate task capability with trust, scoring the two as though they were the same thing.

A human study with thirty participants shows why the principles have to be optimized jointly. Participants' trust declined sharply after an intervention was misaligned, even when the outcome was correct. And they preferred sleep-time assistance, even when it was imperfect, because it preserved their ongoing focus.

Running through all of it is one idea: useful work, compute allocation and user trust cannot be scored separately. A correct result delivered at the wrong moment still costs trust, and a judge that blends capability with trust cannot see that cost. Designing and evaluating proactive agents, the authors conclude, means weighing all three together, with trust treated as something that evolves.