According to the latest reports, Emergence AI has launched a new intelligent web proxy called Agent-E, which boasts a success rate of 73.2%, a 20% improvement over previous versions. This new technology aims to enable autonomous web navigation, allowing AI agents to efficiently complete complex online tasks, from data retrieval and form submissions to ordering the cheapest flights or booking accommodations.
Traditional web proxies often perform inefficiently and are prone to errors when dealing with the complexity and variability of modern web pages. They frequently fail to accurately execute tasks due to an inability to effectively handle noisy and large HTML Document Object Models (DOMs). This inefficiency is a significant barrier to the deployment of autonomous web proxies in practical applications, where reliability and precision are crucial.
Emergence AI's research team has introduced Agent-E, a new web proxy designed to overcome the shortcomings of existing systems. Agent-E employs a layered architecture, dividing the task planning and execution phases into two independent components: the planning agent and the browser navigation agent. Each component focuses on its specific role, thereby enhancing efficiency and performance. The planning agent breaks down user tasks into smaller subtasks and executes them through advanced DOM refinement techniques by the browser navigation agent.
Agent-E's approach includes several innovative steps to effectively manage noisy and large web content. The planning agent breaks down user tasks into smaller subtasks and assigns them to the browser navigation agent. The browser navigation agent uses flexible DOM refinement techniques to select the most relevant DOM representation for each task, reducing noise and focusing on specific task information. Agent-E employs change monitoring to observe state changes during task execution, providing feedback to enhance the agent's performance and accuracy.
Evaluated through the WebVoyager benchmark, Agent-E significantly outperforms previous state-of-the-art web proxy systems. Agent-E achieved a 73.2% success rate, a 20% improvement over previous text-based web proxies and a 16% improvement over multimodal web proxies. On complex websites like Wolfram Alpha, Agent-E's performance improved by 30%. In addition to the success rate, the research team also reported other metrics such as task completion time and error perception. Agent-E averages 150 seconds to successfully complete a task, and 220 seconds for failed tasks. Each task averages 25 large language model calls, highlighting its efficiency and effectiveness.
Emergence AI's research represents a significant advancement in the field of autonomous web navigation. By adopting a layered architecture and advanced DOM management techniques, Agent-E has set a new benchmark for performance and reliability, addressing the inefficiencies of current web proxy systems. The research results indicate that these innovations can be applied to other AI-driven automation fields beyond web automation, providing valuable insights into agent system design principles. Agent-E's success in achieving a 73.2% task completion rate and efficient task execution process underscores its potential to transform web navigation and automation.
Project Entry:
### Key Points:
🌟 Emergence AI launches Agent-E: 73.2% success rate, a 20% improvement
🌟 Agent-E employs layered architecture, DOM management techniques
🌟 Significantly outperforms previous systems in WebVoyager benchmark