Proactive Agent - A new generation of proactive agent interaction paradigm jointly developed by Tsinghua University and Mianbi Intelligence.

What is a Proactive Agent?

Proactive Agent, developed by Tsinghua University in collaboration with teams including Mianbi AI, is a next-generation proactive agent interaction paradigm. It possesses initiative, predicting user needs and taking action without direct commands. Proactive Agent observes the environment and user behavior, infers potential tasks, and autonomously provides assistance. Compared to traditional passive AI agents, Proactive Agent exhibits higher autonomous decision-making capabilities and environmental adaptability, enabling more natural and fluid human-computer interaction experiences in various scenarios. This technological development marks a shift in AI from simple command executors to intelligent collaborative partners with insightful and proactive assistance capabilities.

Main functions of Proactive Agent

• Environmental observation and predictionActively observe the user's environment and behavior to predict the user's needs and intentions.
• Autonomous decision-makingBased on an understanding of the environment and the user's intentions, it makes decisions autonomously without waiting for explicit instructions from the user.
• Task InitiationWhen it is recognized that a user may need help, proactively offer tasks or provide information.
• Context awareness: Understand the context and provide appropriate assistance based on the current situation.
• User InteractionInteract with users, adjust their behavior and predictions based on user feedback, and improve accuracy and user satisfaction.
• Task executionThe user accepts the tasks proposed by the Proactive Agent, the system will execute these tasks, and generate subsequent events as needed.

The technical principle of Proactive Agent

• Environment GymSimulates a specific environment, generates event sequences, maintains the environment state, and provides sandbox conditions for agent interaction.
• Proactive AgentIt receives input from the environment simulator, updates its memory, combines historical interactions and user feedback to predict user intent, and generates tasks.
• User AgentSimulate user behavior, respond to tasks proposed by the Proactive Agent, and decide whether to accept the task.
• Data generation pipeline: Generate data for training and evaluating models based on simulated user activities and responses.
• Reward ModelTrain a model to evaluate the proactive behavior of the Active Agent, simulate human judgment, provide feedback, and optimize the agent's behavior.
• Performance evaluation: Evaluate the performance of the Proactive Agent using metrics such as missing requests, silent responses, correct detection, and error detection, and continuously optimize it.

The project address for Proactive Agent

• GitHub repository:https://github.com/thunlp/ProactiveAgent
• arXiv technical paper:https://arxiv.org/pdf/2410.12361

Application scenarios of Proactive Agent

• Personal AssistantIntelligent scheduling management automatically arranges meetings and reminders based on the user's emails, calendar events, and habits.
• File ManagementAutomatic saving and renaming: When a new file is received, it is automatically saved to the specified location and intelligently renamed according to the file content.
• Life servicesTrip planning: Proactively plan trips and modes of transportation based on users' travel habits and preferences.
• assistive technologyAssistance for visually impaired individuals, providing real-time environmental descriptions, obstacle warnings, and text recognition and reading aloud.
• Office AutomationMeeting scheduling: Automatically schedules meeting times and locations based on team members' schedules and preferences.