Microsoft releases FastContext for code agents

FastContext-1.0, developed by Microsoft, is a lightweight Subagent that helps LLM code agents efficiently navigate code repositories.

Unlike existing approaches where a single model handles both exploration and problem-solving at once, FastContext calls tools such as READ, GLOB, and GREP in parallel and returns the necessary file paths and code ranges as compressed context. This allows the main agent to focus solely on the core task.

Key results and features:

• Improved accuracy: On the SWE-bench Pro benchmark, it recorded a 5.5% performance improvement based on GPT-4o(5.4) and a 5.0% improvement based on GLM-5.1.
• Token savings: In SWE-QA tests based on GPT-4o, it showed token savings of up to 60.3%.
• Efficiency of lightweight models: A lightweight 4B-scale model demonstrated better exploration performance than a 30B-scale model, proving its efficiency.