Shepherd introduces Git-like version control for AI agents

Treating agent trajectories as version-controlled state trees is a major step forward for building reliable, long-horizon AI systems, replacing brittle linear paths with structured search and rollback. Backtracking and rollbacks allow agents to self-correct and recover from mistakes, significantly boosting benchmark performance. Furthermore, reversible traces make complex agent behaviors fully auditable and easier for developers to debug, while forking and branching support parallel experimentation and human-in-the-loop intervention at any execution step.