Sakhanda Wire
NVDA $153.32 -1.81% MSFT $505.84 -2.35% GOOGL $280.31 -1.20% META $621.44 -0.79% AMZN $257.31 -0.17%
← Back to the news

Shepherd lets AI agents fork and replay live execution states

Researchers at Northeastern University and Stanford University have released Shepherd, an open-source Python runtime substrate for managing long-running agent executions. It records every agent-environment interaction as a typed event in a Git-like trace. Unlike ordinary Git history, each commit includes the agent process and filesystem through copy-on-write, allowing a previous live state to be forked and replayed rather than rebuilt from the beginning.

The team reports that Shepherd forks an agent process and its filesystem five times faster than Docker, while replay preserves more than 95% of the prompt cache. The framework defines tasks, effects, runs and workspaces, with permissions specified in task signatures and enforced through native system-call sandboxes. macOS uses Seatbelt, while Linux support uses Landlock in a privileged container.

Reported applications include supervising coding agents, recovering from incorrect tool calls, comparing alternative strategies and generating reinforcement-learning rollouts. In demonstrations, a runtime supervisor raised CooperBench pair-coding pass rates from 28.8% to 54.7%; Tree-RL improved TerminalBench-2 from 34.2% to 39.4%. Shepherd is MIT-licensed, available from PyPI as shepherd-ai and requires Python 3.11 or newer. The project is in early alpha and is not ready for production use.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news