Specialized web agents are significantly outperforming general-purpose LLMs in browser automation tasks, as demonstrated by the Mano-CUA 1.1 model. In the WebRetriever Protocol I benchmark, which evaluates real-world web navigation and form-filling, Mano-CUA 1.1 achieved a score of 41.7 NavEval, surpassing Gemini 2.5 Pro Computer Use (40.9) and Claude 4.5 Computer Use (31.3). Unlike general models that often rely on DOM parsing or accessibility APIs—methods that frequently fail on complex, JavaScript-heavy web applications—Mano-P utilizes a pure vision-driven approach. By analyzing screenshots directly, the model identifies clickable elements through pixels, allowing it to maintain stability even when underlying page structures change. The Mano-P architecture employs a "think-act-verify" loop to handle dynamic UI states, such as dropdowns and validation errors. The training process involves a three-stage pipeline: supervised fine-tuning on GUI interaction data, offline reinforcement learning with logged trajectories, and online RL through live interface interaction. For local deployment, a 4B quantized model can run on Apple M5 Pro hardware at approximately 80 tokens per second. Furthermore, the 72B evaluation model achieved a 58.2% success rate on the OSWorld benchmark, outperforming other specialized models by 13.2 points.