Stanford's Shepherd Brings Git-Style Forking and Rollback to Live AI Agent Runs

Researchers from Stanford University and Northeastern University have released Shepherd, an open-source runtime substrate that records an AI agent's execution like Git and enables forking, reverting, and exact replay of live agent runs. The paper appeared on arXiv in May 2026, and the initial version (v0.1.1) of the GitHub repository was published around July 1. It installs via pip install shepherd-ai and is currently in alpha under active development.