Microsoft's small AI agents get smarter

Microsoft Research is pushing the boundaries of what small AI models can achieve with its latest release: MagenticLite. This experimental agentic application is built to run efficiently across browsers and local file systems within a single workflow, marking a significant step towards capable AI operating directly on user hardware.

The new system combines three core components. MagenticLite itself is a redesigned application, serving as the next iteration of Magentic-UI, optimized with a new harness for smaller models. Powering it are two purpose-built models: MagenticBrain, designed for reasoning and delegation, and Fara1.5, a family of computer-use models focused on browser-based tasks. Fara1.5, building on its predecessor, shows marked improvements in real-world browser interactions, including handling forms and credentialed sites.

This integrated approach underscores a key research bet: that agentic capability hinges more on tool orchestration and action than sheer knowledge. By focusing on these aspects, Microsoft aims to achieve broad agentic task performance with smaller, more cost-effective models.

Doing More With Less

The project's foundation lies in the question of how to make small models genuinely effective at agentic tasks. Microsoft Research's answer involves a holistic redesign across data generation, training objectives, model architecture, and orchestration.

Real-world use cases like form filling and file management informed the development of an evaluation dataset. This scenario-based approach, complementing standard benchmarks, guided iterative improvements to both the models and the execution harness.

The user experience retains key elements from Magentic-UI, such as visibility into the agent's reasoning, user control, and explicit approval at critical junctures. Updates to browser and chat views aim to enhance user comprehension and intervention capabilities.

System Components

Fara1.5: Outperforming its Class

Fara1.5, available in 4B, 9B, and 27B parameter sizes, sets new state-of-the-art results for small computer-use models on benchmarks like Online-Mind2Web. The flagship 9B model nearly doubles the performance of its predecessor, Fara-7B, on web navigation tasks.

Beyond benchmark gains, Fara1.5 offers improved handling of everyday tasks, including form completion and login processes, thanks to advancements in its data generation pipeline. It also features a native action space tuned for long-running tasks, allowing it to store key information and request user input over extended operations.

MagenticBrain: The Orchestrator

As a 14B-parameter orchestration model, MagenticBrain acts as the planner, coder, and delegator. Fine-tuned from Qwen 3 14B, it was trained end-to-end within the MagenticLite harness, ensuring seamless integration between training and inference.

Its design combines multi-step tool-calling with coding capabilities and a specific delegation strategy for computer-use agent (CUA) tasks. This allows MagenticBrain to fluidly reason, code, call tools, and hand off browser tasks to Fara1.5.

The Harness: Optimized for Small Models

The execution harness is central to MagenticLite's efficiency. It employs step-by-step planning and active context management to keep prompts focused and models effective, even with smaller context windows.

Delegation through subagents, where MagenticBrain hands off specialized work to Fara1.5, plays to the strengths of individual models. The harness also preserves human-in-the-loop guarantees and operates within a sandboxed environment for security.

MagenticLite, MagenticBrain, and Fara1.5 are available as research releases on GitHub and Microsoft Foundry, inviting community experimentation and feedback.