Shepherd can checkpoint a live agent process, not just files, and rewind it 5× faster than Docker

Ask any engineer who runs long-horizon coding agents what they really want, and they’ll tell you: an undo button that actually works. Not a file diff. A full state rewind. When an agent at step ten misreads a traceback and corrupts a working file, you’re stuck with two bad options — patch forward and bloat your context window, or restart from zero and pray the non-deterministic run reproduces anything useful. Neither is cheap.

Researchers at Northeastern and Stanford have built Shepherd, a Python runtime substrate that treats an agent’s entire execution as a Git-like trace. But here’s the difference that matters: a commit captures the agent process and the filesystem together, copy-on-write. Forking from a past commit restores live state — the edited files, the running dev server, the warm prompt cache. The team reports forks that are 5× faster than Docker, with over 95% prompt-cache reuse on replay because the prefix through the branch point hasn’t changed.

On top of this forkable substrate, they’ve demonstrated what a meta-agent can do when it can observe and intervene. A live supervisor watching a coding agent’s trace lifted pair-coding pass rates on CooperBench from 28.8% to 54.7%. That’s not incremental. It’s the difference between a tool you supervise nervously and one you might actually trust to run while you get coffee. Branching exploration over candidate strategies also beat baselines by up to 11 points while cutting wall-clock time by up to 58%.

Shepherd is MIT-licensed and pip-installable now, though it’s early alpha — not something you’d wire into a production pipeline tomorrow. It requires Python 3.11+ and enforces permissions at the OS level via Seatbelt on macOS or Landlock on Linux. The design formalizes agent interactions as typed effects with declared permissions, meaning a May[GitRepo, ReadOnly] binding compiles to a syscall jail for that run. That’s the kind of guardrail that makes long-running autonomous agents feel less like a gamble. The real shift here isn’t just faster checkpointing — it’s that agent state becomes a first-class object you can branch, compare, and roll back. Once that’s cheap, the whole way we supervise these systems changes.