MagenticLite is here: A full-stack agentic experience powered by Small Models
Microsoft Research introduces MagenticLite, an agentic framework powered by two new small, open-weight models: Magentic Orchestrator for planning and coding, and Fara-1.5 for browser automation. The talk details the novel synthetic data generation techniques and training strategies used to achieve state-of-the-art performance in small models, enabling them to compete with much larger ones.
Microsoft Research’s AI Frontiers Lab has introduced MagenticLite, the next generation of the Magentic UI, an agentic system designed to be efficient, capable, and no longer reliant on frontier-scale models. The release is an end-to-end integration across the agentic stack, including the MagenticLite user experience and two new small language models that power it: Magentic Orchestrator and Fara-1.5.
The team’s development philosophy moved beyond standard benchmarks, focusing instead on “hero use cases”—real-world tasks—to build custom evaluations. This approach created a flywheel of iterative improvements across the entire stack, from data generation to user experience.
MagenticLite: The Agentic Experience
MagenticLite is a lightweight and fast agentic application with a refreshed UX based on community feedback. It operates across both a web browser (running in a secure virtual machine) and the user’s local file system, allowing it to perform a wide range of tasks:
- Managing and processing local files.
- Searching the web to gather information.
- Generating documents and code.
- Filling out online forms and sending emails.
A key feature is its collaborative nature, where the agent can perform long-running tasks autonomously but will notify the user when their attention or intervention, such as logging into an account, is required.
Magentic Orchestrator: The Brain
The Magentic Orchestrator model acts as the central brain of the system, functioning as a planner, coder, and delegator. It deconstructs complex, high-level user requests (e.g., “Book me a dentist appointment Tuesday afternoon, and add it to my calendar”) into a concrete, multi-step plan. It then determines the right tool or sub-agent for each step, writes Python code when a specific tool isn’t available, and handles error recovery.
The main innovation is achieving this complex reasoning capability within a small model using a simple supervised fine-tuning (SFT) approach. The key to its success lies in the training data mix, which blends two complementary data styles:
- Standard Tool Calling Data: Generated using an in-house pipeline called
vulcan, this data teaches the model realistic, multi-turn tool compositions. Tools are defined as Python objects, and trajectories are modeled as paths through a graph of these tools, moving beyond simple, isolated tool calls. - Terminal Data: This data consists of “cowork style” trajectories where the agent thinks and executes actions step-by-step in a terminal. Generated via a four-stage pipeline, this teaches the model to write short scripts or code snippets when a pre-defined tool doesn’t fit the task.
By mixing these two styles, the Orchestrator learns the critical skill of deciding when to use a tool versus when to write code, resulting in a small, open-weight model that competes with much larger general-purpose models.
Fara-1.5: The Browser Agent
Fara-1.5 is a family of state-of-the-art computer-use agent models (4B, 9B, and 27B) designed for browser automation. The 9B model nearly doubles the performance of its predecessor on the Mind2Web benchmark, achieving a 65% success rate.
Its operational loop is straightforward:
- Capture a screenshot of the browser.
- Process the screenshot, past actions, and the user prompt as context.
- Predict and execute a native action on the computer (e.g., clicking coordinates, typing).
- Observe the result and continue the loop.
The model’s action space includes tools for managing its own context, such as memorizing facts for long tasks, and the ability to ask the user for approval or clarification.
The core of Fara-1.5’s success is FaraGen 2.0, a scalable synthetic data engine built on the principle of using zero human interaction data. Its key components are:
- Environments: A mix of live, popular websites and fully synthetic websites. These synthetic domains, generated by a coding agent, allow Fara to be trained on scenarios (like booking a flight) that are difficult or impossible to automate on the live web.
- Solver System: A powerful teacher agent and a user simulator work together to generate successful task completion trajectories.
- Verifiers: A suite of verifiers judges the generated trajectories on three criteria:
- Correctness: Did the agent complete the task correctly?
- Efficiency: Did the agent take the most straightforward path?
- User Interaction: Did the agent avoid making up information or performing irreversible actions without approval?
Future Directions
Work on computer use agents is ongoing. Future plans for Fara include enabling always-on capabilities for monitoring the web, expanding its operational environments to include Windows and Linux, and integrating terminal use alongside visual manipulation for greater efficiency.