Shepherd

Shepherd is a Python runtime substrate that represents agent execution as reversible, Git-like traces. It allows meta-agents to observe, fork, replay, and revert agent runs, enabling higher-order agents to supervise and optimize other agents. The tool structures every model action, tool call, and environment change as a reversible event, allowing any past state to be reverted significantly faster than traditional containerization approaches and supporting efficient replay with high key-value cache reuse.

Shepher is developed by shepherd-agents and licensed under the MIT license. The substrate is designed to make agent execution a first-class object that meta-agents can easily inspect and transform. It enables use cases such as runtime supervision to prevent agent conflicts, counterfactual optimization to repair failed agent workflows, and training meta-agents that improve credit assignment in reinforcement learning scenarios. The tool couples agents and environments using copy-on-write fork operations for performance.