Shepherd: Stanford’s Git-Like Rollback System for AI Agents
Stanford researchers have created a Git analogue for AI agents.
It’s called Shepherd.
When an AI agent makes a mistake while performing a task, it doesn’t need to start over. Instead, the system rolls back to the last correct step and continues working from there.
Shepherd saves each execution stage as a separate commit—along with files and the state of running processes. You can instantly return to any point in the execution with a single call. Up to 95% of the computations performed in previous steps are reused, so nothing has to be recomputed twice.
On top of everything runs a meta-agent that automatically performs a rollback as soon as it detects an error. No human involvement is required.