🚀 BrowseMaster: Scalable Web Browsing via Tool-Augmented Programmatic Agent Pair

BrowseMaster is a next-generation LLM-based browsing agent that addresses the shortcomings of traditional search and existing agents by uniting broad, code-driven search with deep, strategic reasoning.

Its Planner–Executor framework delivers scalable, coherent, and efficient information seeking, achieving strong results on both BrowseComp and BrowseComp-ZH benchmarks.

🛠️ Framework

🔑 Key Features

• 🧠 Planner–Executor Framework – Separates large-scale, code-driven search from high-level reasoning for cleaner context and deeper thinking.
• 📜 Code-Based Search Primitives – Execute batch_search, check_condition, generate_keywords operations over thousands of pages with minimal context overhead.
• 🎯 Confidence-Guided Replanning – Dynamically reset strategy when confidence is low to avoid premature convergence.

📖 Documentation

Prompts and Examples

• Detailed prompts for both planner and executor are available in the prompts/ directory
• Sample responses from BrowseComp evaluations can be found in logs/example.jsonl

Getting Started

Prerequisites

• For sandbox setup and code execution environment, please refer to mcp_sandbox
• Note that we use locally deployed DeepSeek series models, instead of api.

Configuration

1. Configure model endpoints in configs/model_config.yaml:
   • DeepSeek-R1-0528 model URL
   • DeepSeek-R1 model URL
   • ToolBox URL

Running BrowseMaster

1. Ensure all prerequisites are met (environment, toolbox, and configuration)
2. For single query inference, execute:

python agent.py

📊 Experiments

🚀 Scaling Search Calls & Computation Efficiency

🌐 Breadth of Domains Explored