Shepherd: Open-Source Python Runtime Letting Meta-Agents Fork, Replay, and Revert Any Agent Run
Researchers from Northeastern University and Stanford University have released Shepherd, an open-source Python runtime substrate that records agent runs as Git-like traces of typed events, enabling forking and replaying of any past state. It achieves 5× faster forks than Docker and over 95% prompt-cache reuse on replay, and is available in early alpha via pip.
Shepherd is a Python runtime substrate that records an agent's execution as a first-class object, with each agent-environment interaction becoming a typed event in a Git-like trace. Unlike Git, each commit covers the agent process and filesystem together using copy-on-write, so a branch carries live state, not just files. This allows forking from any earlier commit, enabling a meta-agent to observe a trace and intervene before a bad write commits. The research team reports forks 5× faster than Docker and over 95% prompt-cache reuse on replay. Demonstrations show a live supervisor raising pair-coding pass rates on CooperBench from 28.8% to 54.7%, branching exploration beating baselines by up to 11 points with 58% less wall-clock time, and Tree-RL training improving TerminalBench-2 from 34.2% to 39.4%. Shepherd requires Python 3.11+ and supports OS-level grant enforcement on macOS (Seatbelt) and Linux (Landlock). It is MIT-licensed and installable via pip install shepherd-ai.
- Abbreviations
- RL = Reinforcement Learning — обучение с подкреплением
- MIT = Massachusetts Institute of Technology — Массачусетский технологический институт
Source: MarkTechPost —
original