Shepherd Launches Alpha Runtime That Makes AI Agent Executions Reversible With Git-Like Forking and Sandboxed Permissions

Shepherd launches an alpha runtime that makes AI agent executions reversible using Git-like forking, delivering 5x faster copy-on-write performance than Docker commit, 95% KV-cache reuse on replay, and OS-level sandboxed permissions that block unauthorized writes before they happen.

• Shepherd is a runtime substrate that converts agent executions into reversible, Git-like traces, enabling meta-agents to observe, fork, replay, and revert any run with copy-on-write forking that is approximately 5x faster than Docker commit and achieves roughly 95% KV-cache reuse on replay.
• The framework enforces a signature-based permission system where tasks declare read-only or read-write grants per repository binding, with OS-level syscall enforcement via macOS Seatbelt and Linux Landlock, ensuring unauthorized writes are blocked before execution rather than caught at a merge gate.
• Currently in early alpha and requiring Python 3.11+, Shepherd supports sandboxed agent workflows where task outputs are held as reviewable proposals that users can accept, merge, or discard, with full documentation available at docs.shepherd-agents.ai and reproducible paper experiments hosted in a companion repository.