Shepherd enables AI agents to fork, replay, and recover from errors five times faster than Docker with 95% prompt-cache reuse.
On 8 August 2026, researchers from Northeastern and Stanford universities introduced Shepherd, an open-source Python runtime treating an AI agent’s execution history as a versioned, Git-like object. Released under the MIT Licence (pip install shepherd-ai), it is available in early alpha, requires Python 3.11+, and enforces OS-level sandboxing via macOS Seatbelt and Linux Landlock.
Long-horizon coding and DevOps agents accumulate live state, running processes, edited files, packages, dev servers, and prompt caches, untracked by simple transcripts. When an agent errors at the latest step, patching forward inflates context costs, whilst restarting from the beginning loses progress and invalidates caches. Shepherd solves this by recording every agent–environment interaction as a typed event, capturing both the process and filesystem state in copy-on-write commits. Meta-agents can thus rewind, fork, or replay from any previous turn.
The runtime forks process and filesystem states 5 times faster than Docker, achieving over 95 per cent prompt-cache reuse during replay. For irreversible real-world actions, Shepherd enforces syscall-level refusal rather than rollback.
Empirical benchmarks show substantial gains: a supervisory meta-agent increased CooperBench pair-coding pass rates from 28.8% to 54.7%. Branching exploration beat baselines by up to 11 percentage points whilst cutting wall-clock execution times by up to 58%, and reinforcement learning rollouts improved TerminalBench-2 performance from 34.2% to 39.4%.
