Shadow-Frog: Coding Agents that Dream and Discover via Active Exploration
Microsoft GitHub has released Shadow-Frog, an agentic discovery system that builds a codebase memory for coding agents through active exploration rather than passive recording. During 'discovery sessions,' agents run experiments on underexplored parts of the codebase, capturing learnings in a structured shadow knowledge base. Shadow-Frog improves retrieval accuracy to 97.6%, finds synthetic bugs 25.4 percentage points better than baselines, and flags 88% of real bugs, demonstrating its ability to acquire tacit knowledge by 'learning by doing.'
This system allows coding agents to proactively learn and build an understanding of a codebase, leading to better bug detection, knowledge retrieval, and even feature ideation, moving beyond passive memory systems.
Learn one new AI thing every day.
Daily Deck sends you seven plain-English cards like this every morning. Free.
Start free