WebAgent (WebWalker & WebDancer) - AI Agent Frameworks Tool

Overview

WebAgent is an open-source autonomous agent framework from Alibaba Group focused on information-seeking across the live web. It packages two complementary components: WebWalker, a benchmark (presented as an ACL2025 benchmark) that evaluates large language models on long-horizon web traversal, and WebDancer, a native agentic search reasoning model described in a project preprint. The project adopts the ReAct (Reason+Act) interaction pattern to interleave reasoning and web actions, enabling models to plan multi-step sequences of queries and navigation to gather and synthesize information. The codebase emphasizes a four-stage training paradigm that the authors describe as combining supervised fine-tuning, reward modeling, and reinforcement learning to improve long-horizon planning and decision-making on web tasks. WebAgent is intended for researchers and engineers building autonomous search agents, for benchmarking LLM capabilities on realistic web tasks, and for reproducing or extending the training strategies shown in the preprint and ACL2025 benchmark. According to the GitHub repository and accompanying papers, the project targets multi-step web traversal scenarios where agents must act, observe, and revise plans to complete complex information-seeking objectives.

Installation

Install via pip:

git clone https://github.com/Alibaba-NLP/WebAgent.git

cd WebAgent

pip install -r requirements.txt

Key Features

• WebWalker benchmark: evaluates LLMs on long-horizon, multi-step web traversal tasks (ACL2025).
• WebDancer model: native agentic search reasoning model described in the project preprint.
• ReAct integration: interleaves reasoning and actions for stepwise web decision making.
• Four-stage training paradigm: supervised fine-tuning, reward modeling, and reinforcement learning.
• Designed for autonomous search: supports iterative query, observation, and plan revision workflows.

Community

WebAgent is published as an open-source GitHub project under the Alibaba-NLP organization; the repository links the ACL2025 benchmark (WebWalker) and a preprint describing WebDancer. According to the project page, development and evaluation are research-driven and intended for community reproduction and extension. Users and contributors can engage via the repository’s issues and pull requests to report bugs, request features, or discuss experimental setups; the codebase and papers are the primary entry points for collaboration and citation.

Key Information

• Category: Agent Frameworks
• Type: AI Agent Frameworks Tool