TLDR: BrowseMaster is a novel AI framework that enhances web browsing for large language models by using a planner-executor agent pair. The planner focuses on high-level reasoning and strategy, while the executor efficiently conducts broad, programmatic searches using code-driven tools. This design allows BrowseMaster to overcome limitations in search breadth and reasoning depth, outperforming existing AI agents on complex information-seeking tasks across various benchmarks.
In the vast and ever-expanding digital world, finding information effectively means striking a balance between searching widely and thinking deeply. Traditional large language model (LLM)-based agents often struggle with this. They tend to search slowly, one step at a time, which limits how many sources they can check. Also, the raw, unorganized information they get from the web can disrupt their multi-step thinking process, making it hard to reason deeply.
To tackle these challenges, researchers have introduced BrowseMaster, a new framework designed for scalable web browsing. It’s built around a clever system of two cooperating agents: a ‘planner’ and an ‘executor’. This division of labor is key to its success, allowing it to maintain clear, long-term reasoning while exploring a wide range of information systematically.
The Planner acts as the long-term strategist. Its main job is to understand the user’s request, identify important constraints, and create a search strategy. It breaks down complex tasks into smaller, manageable sub-tasks and hands them over to the executor. Crucially, the planner only works with organized information provided by the executor, which keeps its thinking process clean and prevents it from getting bogged down by messy web content. It even has a ‘confidence-guided replanning’ feature, allowing it to rethink its strategy if it’s not confident in its current path, ensuring thorough exploration.
The Executor is the scalable search engine. It’s responsible for efficiently gathering as much accurate and relevant information as possible. Instead of relying on natural language commands for every action, the executor uses programmatic tools. This means it can write and execute code to perform actions like searching, parsing web pages, and checking conditions. This code-driven approach allows it to perform many operations at once, like running multiple web searches in parallel, which is much faster than traditional methods.
BrowseMaster uses a set of standardized programming ‘primitives’ to make its operations efficient. These include `generate_keywords` to create diverse search terms, `batch_search` to execute multiple searches simultaneously, and `check_condition` to filter and evaluate web content programmatically. These tools help the executor handle large volumes of information quickly and accurately.
The framework also includes two essential tools: a ‘web search’ tool and a ‘web parse’ tool. The web search tool uses a search engine to find relevant pages, providing summaries, titles, URLs, and related queries. The web parse tool can then delve deeper into selected pages, extracting main content and identifying related links. For scientific papers, it even attempts to fetch HTML versions or downloads PDFs for detailed analysis.
A significant innovation in BrowseMaster is its ‘stateful code execution sandbox’. Unlike typical sandboxes that forget information after each execution, BrowseMaster’s environment remembers variables and functions defined in previous steps. This is similar to how a Jupyter Notebook works, giving the agents much more flexibility in their coding and multi-step tool use.
Extensive experiments have shown BrowseMaster’s impressive capabilities. It consistently outperforms both open-source and proprietary agents on challenging benchmarks like BrowseComp (English and Chinese versions), xBench-DeepResearch, GAIA, and WebWalkerQA. For instance, it achieved a 30.0% score on BrowseComp-en, making it the first open-source agent to reach this milestone. It even surpassed OpenAI’s DeepResearch on BrowseComp-zh by 4%.
The research highlights that scaling search calls and computational resources are crucial for enhancing agent performance. BrowseMaster’s programmatic tool use significantly boosts search efficiency and enables broader exploration, allowing it to navigate thousands of pages and reason effectively over diverse search cues. The number of interactions between the planner and executor also reveals the complexity of tasks, with more complex tasks requiring more back-and-forth communication and confidence-guided retries.
Also Read:
- HierSearch: A New Framework for Enterprise Deep Search Across Local and Web Data
- Understanding Advanced AI Search Agents
In essence, BrowseMaster represents a significant step forward in automated information seeking. By combining strategic reasoning with efficient, code-driven web exploration, it sets a new standard for how AI agents can tackle complex, real-world information challenges. You can find more details about this innovative framework in the full research paper: BrowseMaster: Towards Scalable Web Browsing via Tool-Augmented Programmatic Agent Pair.
