Microsoft Unveils Browser Automation Agent Fara1.5-27B

Fara1.5-27B, developed by Microsoft Research AI Frontiers, is a multimodal computer use agent (CUA) that autonomously operates web browsers. It was supervised fine-tuned based on Qwen3.5-27B, and it perceives browser state solely through screenshots, without relying on the DOM or accessibility tree.

The model performs tasks through structured tool calls such as click, type, scroll, visit URL, and web search, predicting pixel-level precise arguments like click coordinates.

Key use cases:

• Automating repetitive web tasks such as filling out forms, shopping, and travel/restaurant reservations
• Serving as a grounding model for other agents that require pixel-accurate action prediction

The training data was synthesized using their own multi-agent pipeline FaraGen1.5, and deployment alongside MagenticLite is recommended. Limitations are explicitly noted, including English-only training data, vulnerability to prompt injection, and error accumulation across multiple steps.