This project provides a framework for building LLM-powered proactive agents that anticipate user needs and offer assistance without explicit requests. It targets developers and researchers interested in creating more intuitive and helpful AI assistants for coding, writing, and daily life scenarios. The core benefit is enabling agents to act proactively, enhancing user experience and productivity.

How It Works

The system employs a data generation pipeline that includes "Environment Gym" for simulating user activities and "Activity Watcher" for collecting real-world traces. This data is used to train a Proactive Agent and a Reward Model. The Reward Model, achieving a 0.918 F1 score, evaluates the agent's proactive suggestions, aiming to align them with human preferences. The pipeline allows for dynamic data generation, incorporating user feedback to refine future agent behavior.

Quick Start & Requirements

• Installation: Clone the repository, create a conda environment (conda create -n activeagent python=3.10), activate it (conda activate activeagent), and install dependencies (pip install -r requirements.txt).
• Activity Watcher: Download the main app from the official website and install browser extensions (Chrome/Edge) or VSCode extension. Verify installation by checking http://localhost:5600/#/timeline.
• Configuration: Copy example_config.toml to private.toml and update API keys and model settings.
• Usage: Navigate to the ./agent folder and follow instructions for running the agent. Proposals are displayed as toasts, with options to accept, reject, or ignore.
• Dependencies: Python 3.10, conda, and specific libraries listed in requirements.txt.

Highlighted Details

• The Reward Model achieves a 0.918 F1 score on its test set, indicating strong performance in evaluating proactive agent suggestions.
• Benchmarks show the custom "ours" model outperforming several leading LLMs (GPT-4o, LLaMA-3.1) in metrics like F1 score and Accuracy when fine-tuned for proactive tasks.
• The project includes a comprehensive data collection and generation pipeline, along with annotated datasets for coding, writing, and daily life scenarios.
• An interactive demo is available to showcase the agent's capabilities.

Maintenance & Community

The project is from THUNLP and has been accepted by ICLR 2025. Model releases for Proactive Agent and Reward Agent are available. Further improvements to data quality and scenario coverage are planned.

Licensing & Compatibility

Distributed under the Apache License 2.0. This license is permissive and generally compatible with commercial use and closed-source linking.

Limitations & Caveats

The current data focus is limited to coding, writing, and daily life scenarios. The Activity Watcher extension is not tested on Safari. The project is under active development, with some features marked "TO BE UPDATE".