A team at Northeastern University and Stanford has released an open-source framework that gives autonomous agents something software engineers have had for two decades: a version history they can roll back. Shepherd records every model decision, tool call, agent state and relevant slice of the environment as a structured commit, so a supervisory agent can jump to any earlier point, rewind a mistake and fork a different attempt from there.
The work, led by Northeastern assistant professor Weiyan Shi with Stanford's Derek Chong, Simon Yu and Christopher Manning, is scheduled for presentation at NeurIPS in December. Shi described the division of labour with a farmyard metaphor: the worker agents are sheep wandering around, and the meta-agent is the shepherd trying to bring them back on track.
The design borrows directly from Git. Instead of handing operators a raw transcript after a long run finishes, Shepherd stores execution as a commit graph with copy-on-write environment forks, so replaying a corrected path does not require re-running the whole task. A supervisor can observe a worker without interrupting it, revert a step, or branch off to try an alternative fix.
The safety mechanism is a sandbox that intercepts consequential actions. When a worker agent drafts and sends an email to a client, the action appears to succeed from the worker's point of view. In reality the outbound call is held as a proposal inside Shepherd until a meta-agent or a human approves it, which converts an irreversible external effect into a reviewable one.
The reported results are the interesting part. A supervisor meta-agent watching parallel coding workers caught merge conflicts and divergent edits mid-run, raising pair-coding pass rates on CooperBench from 28.8 percent to 54.7 percent. An optimisation meta-agent that replayed only the downstream commits affected by each edit beat MetaHarness by 12.8 percent on Terminal-Bench 2.0 while cutting wall-clock time by 58 percent. A third meta-agent chose which commits to fork during long-horizon reinforcement learning rollouts and doubled GRPO's uplift on the same benchmark.
Those numbers come from the group's own paper and its own harness, which is the standard caveat for agent infrastructure research. What is more persuasive is the shape of the contribution: an architectural fix that addresses supervision, debugging and token cost at once, rather than another prompt-level guardrail that a determined model can talk its way past.
The code is public under an MIT licence and installable from PyPI, but Chong said the project is still moving from alpha toward production readiness, and the published version is an early release. Shepherd sits inside the Agent Sciences Research Initiative, a joint effort across Northeastern, Stanford and UC Berkeley studying agent reliability and evaluation.
What this means: the industry has spent a year trying to make agents safer by asking models to police themselves. Shepherd suggests the more durable answer is to treat execution traces as versioned data and put the review layer outside the model, which is also, not coincidentally, how every other kind of automation that touches production got reliable.
Comments (0)
Log in to join the discussion
Log InNo comments yet