MagenticLite

MagenticLite is an agentic application from Microsoft Research AI Frontiers — the next generation of Magentic-UI, redesigned around a pair of small models rather than frontier ones. The app pairs a refreshed chat and browser view with an agent harness rebuilt to actively curate context, plan step by step, and delegate browser tasks to a specialist sub-agent. Browser sessions and code execution run inside Quicksand, the project’s QEMU sandbox, so data stays on the user’s machine. Users can watch the agent’s reasoning, take direct control at any point, and approve critical actions like logins and submissions.

MagenticLite tests a clear research hypothesis: agentic capability lives in tool orchestration and action, not in raw model knowledge. By co-designing the app, the harness, and a pair of small models — MagenticBrain as the orchestrator and Fara1.5 as the computer-use specialist — Microsoft Research delivers strong agentic performance at a fraction of the cost of a frontier-model stack. The result targets developers, researchers, and power users who want autonomy and transparency on commodity hardware, and serves as Microsoft’s reference architecture for the small-model agentic stack alongside Magentic-One and Magentic-UI.