Researchers from Shanghai Jiao Tong University and Tencent have developed a proactive AI agent named ProAct. Unlike standard reactive systems, ProAct utilizes downtime between user interactions to predict likely follow-up questions by analyzing past conversations and user data. The system then prepares information in advance, aiming to reduce overall conversation turns and follow-up requests. In simulated testing across various domains, ProAct reportedly anticipated significantly more user needs and reduced instances of AI-generated false information, or hallucinations, by 28.1%.
Researchers at Shanghai Jiao Tong University and Chinese technology conglomerate Tencent have developed an AI agent called ProAct designed to predict user needs before they are asked. The system works during idle time between messages to review past conversations and saved information, preparing useful responses in advance.
“While AI agents demonstrate remarkable capabilities in reasoning and tool use, they remain fundamentally reactive,” the researchers wrote in their paper. They stated the new paradigm addresses a critical opportunity often wasted between interactions. The ProAct system operates through stages of Future-State Prediction and Idle-Time Acquisition to decide what to research and when to present information.
In benchmark testing involving 200 simulations, ProAct reduced conversation turns by 14.8% and cut follow-up requests by 11.7%. The system anticipated 703 predictable user needs compared to just 32 for an earlier proactive system in a comparison using a benchmark called ProActEval. The research emerges alongside the broader industry trend of autonomous AI agents like OpenClaw and Hermes Agent handling more independent tasks.
Researchers acknowledged limitations, including that in 3% of cases the system worsened responses with irrelevant information. They noted any real-world version would require robust privacy protections due to its constant analysis of conversations. Separate research this month has warned that such autonomous agents can sometimes complete tasks without fully understanding consequences, as discussed by lead author Erfan Shayegani of UC Riverside.
