Researchers at Shanghai Jiao Tong University and Chinese technology conglomerate Tencent claim to have built an AI agent that uses the quiet time between conversations to predict what users may ask next—and prepare answers before they ask.
The system, called ProAct, works differently from most AI agents that wait for users to ask a question before responding. Instead, ProAct uses downtime between messages to review past conversations and saved user information, then prepares useful information in the background before the next question arrives.
"While AI agents demonstrate remarkable capabilities in reasoning and tool use, they remain fundamentally reactive: They compute responses only after explicit user prompts," the researchers wrote. "This paradigm ignores a critical opportunity: The idle time between interactions is largely wasted, leaving agents unable to prepare for future user needs."
The system works in multiple stages. The first, called Future-State Prediction, predicts likely follow-up questions by analyzing past conversations, user preferences, and missing information. The second stage, called Idle-Time Acquisition, decides which of those predictions are worth researching based on relevance, timing, and how useful the new information might be.
A separate system then decides whether to present the prepared information, save it for later, or store it until needed, creating a "closed-loop" system designed to anticipate and respond to user needs.
"After each foreground interaction, the agent updates its memory, predicts possible future needs, allocates idle-time computation to valuable candidates, and decides how the resulting preparation should be handled," the researchers wrote. "This formulation ties prediction, acquisition, and delivery to a single policy, rather than treating idle-time compute as unconstrained background search."
According to the researchers, ProAct was tested in 200 simulations across 40 domains, including financial planning, software release management, and cybersecurity. The system reduced conversation turns by 14.8% and cut follow-up requests by 11.7%. In a comparison using a benchmark called ProActEval, ProAct anticipated 703 predictable user needs versus 32 for an earlier system. The researchers also reported a 28.1% reduction in hallucinations.
The research comes as autonomous AI agents spread across the tech industry, with projects such as OpenClaw and Hermes Agent delivering persistent AI assistants capable of handling longer, more independent tasks—such as coding, scheduling, research, and workflow automation—with less direct human input.
The study also arrives as separate researchers earlier this month warned that AI agents may complete dangerous tasks without fully understanding the consequences. "Like Mr. Magoo, these agents march forward toward a goal without fully understanding the consequences of their actions," said lead author Erfan Shayegani, a UC Riverside doctoral student, in a statement. "These agents can be extremely useful, but we need safeguards because they can sometimes prioritize achieving the goal over understanding the bigger picture."
Researchers acknowledged that the ProAct study had several limitations, including that in 3% of cases the system made responses worse by surfacing irrelevant information. The paper also noted that any real-world deployment would require privacy protections, given that the system continuously analyzes conversations and stores user data.
"Our budget analysis further shows that larger Idle-Time Acquisition budgets raise active-token cost and yield diminishing returns," the researchers wrote, "so proactive computation is an operating-point trade-off rather than something to maximize."
The researchers also noted that the experiments did not involve real users, which remains a key limitation of the study's findings.
Why it matters
ProAct's architecture shifts when AI computation happens—moving work into idle periods rather than only after a prompt—which has implications for how latency and token costs are managed in deployed agent systems.
The acknowledged 3% degradation rate, where pre-fetched information made responses worse, highlights a concrete reliability trade-off that developers and enterprises would need to account for before deploying proactive agents in sensitive domains like financial planning or cybersecurity.
The researchers' note that larger idle-time budgets yield diminishing returns frames proactive computation as a tunable operating parameter, not a default improvement—meaning real-world implementations would require deliberate configuration rather than simply enabling the feature.
