A specialized web agent scored 41.7 on WebRetriever while GPT and Claude failed the same form-filling task

Описание и применение

Специализированный агент Mano‑CUA достиг 41,7 баллов на WebRetriever, превзойдя GPT и Claude в сложных задачах заполнения форм и навигации по веб‑страницам. Он использует визуальное распознавание и цикл think‑act‑verify, работая полностью локально на Mac с Apple silicon.

Задача

General-purpose LLMs struggle with complex browser automation tasks such as multi-step form filling, dropdown selection, and navigation through paginated results.

Результат

Mano-CUA achieved a 41.7 NavEval score on WebRetriever Protocol I, surpassing Gemini 2.5 Pro (40.9) and Claude 4.5 (31.3). It reliably handles multi-step form filling, dropdowns, and dynamic page changes through a vision‑driven think‑act‑verify loop.

Рабочий процесс

• Install Mano-CUA via Homebrew on Apple silicon
• Run the CLI with the --local flag to keep all data on-device
• Provide a natural language instruction (e.g., "fill out the application form on example.com")
• The agent captures screenshots, identifies clickable elements visually, and executes actions
• After each action it verifies the result from the new screenshot and retries if necessary

Доказательства

Mano-CUA 1.1 scored 41.7 NavEval. Gemini 2.5 Pro Computer Use got 40.9. Claude 4.5 Computer Use landed at 31.3.

Источник

The 4B quantized model runs on Apple M5 Pro hardware at about 80 tokens per second decode speed.

Источник

Mano-CUA’s think‑act‑verify loop forces it to check whether each action actually worked before moving to the next step.

Источник