Mango: Multi-Agent Web Navigation via Global-View Optimization

A team of researchers has introduced Mango, a novel multi-agent framework designed to solve a fundamental flaw in current web navigation AI. Existing agents typically start exploring from a website's root URL, forcing them to crawl through deep, hierarchical structures inefficiently. This often leads to getting stuck in navigation traps, exploring irrelevant sections, or failing to find target information within limited step budgets. Mango addresses this by giving agents a "global view" of a website's structure, allowing them to dynamically select optimal starting points instead of always beginning at the homepage.

The system formulates the challenge of choosing where to start as a multi-armed bandit problem, a classic reinforcement learning scenario. It employs Thompson Sampling to intelligently allocate its navigation budget (like a limited number of clicks) across various candidate URLs on a site. Furthermore, Mango incorporates an episodic memory component, enabling it to store and learn from the history of its navigation attempts on similar sites. This combination allows it to learn efficient pathways over time.

In rigorous testing, Mango demonstrated significant performance gains. On the WebVoyager benchmark using GPT-5-mini as its reasoning engine, it achieved a 63.6% success rate, outperforming the previous best baseline by 7.3%. Its improvement was even more pronounced on the WebWalkerQA benchmark, where it attained a 52.5% success rate—a substantial 26.8% absolute improvement over the best existing method. The researchers also showed Mango's architecture is model-agnostic, working effectively with both open-source and closed-source AI models as its backbone. The project's code and data have been released as open-source, paving the way for broader development in autonomous web agents.