Researchers from Shanghai Jiao Tong University and Tencent claim they have built an AI agent that predicts what users might ask next and prepares answers in advance. The system, named ProAct, uses the quiet time between messages to review past conversations and user data. It then gathers useful information before the user actually asks a question.
The researchers point out that most AI agents only respond after a user gives a prompt. They call this reactive approach a missed chance. Speaking about their work, they said the idle time between chats is largely wasted, so agents cannot get ready for future needs.
How ProAct works
ProAct operates in distinct stages. First, it predicts likely follow-up questions by analyzing past conversations, user preferences, and gaps in information. The system then decides which predictions to research based on relevance, timing, and potential usefulness. A separate mechanism decides whether to show the prepared answer, save it for later, or just store it. This creates what they call a closed-loop system that can respond proactively.
According to the study, this avoids computing everything on the fly. They wrote that after each interaction, the agent updates its memory, predicts possible needs, allocates idle-time computation, and decides how to handle the results. This ties prediction, acquisition, and delivery together rather than treating idle time as random background search.
Test results and limitations
The researchers tested ProAct in 200 simulations across 40 domains, including financial planning, software release management, and cybersecurity. They observed a 14.8% reduction in conversation turns and an 11.7% decrease in follow-up requests. Using a benchmark called ProActEval, ProAct anticipated 703 predictable user needs compared to just 32 for an earlier system. There was also a 28.1% drop in hallucinations.
Still, the study had limitations. In about 3% of cases, the system made responses worse by introducing irrelevant information. The paper also noted that real-world deployment would need strong privacy protections, since the system constantly analyzes conversations and stores user data. Budget analysis showed that larger idle-time computation budgets raise active-token costs and lead to diminishing returns.
The broader context
This research comes as autonomous AI agents grow across the tech industry. Projects like OpenClaw and Hermes Agent already provide persistent assistants that handle longer, independent tasks such as coding, scheduling, and workflow automation with less human input.
But there are also warnings. Separate researchers recently cautioned that AI agents might complete dangerous tasks without understanding the consequences. One lead author said these agents can be extremely useful but sometimes prioritize achieving the goal over understanding the bigger picture. They compared the behavior to a cartoon character marching forward blindly.
In summary, ProAct shows promise for proactive AI systems, but it also highlights trade-offs in cost, privacy, and reliability. The work suggests that making AI less “reactive” is possible, but perhaps it comes with its own set of complications.
