Show HN: Teleport-env – <500ms stateful rollbacks for AI agents via CRIU (github.com)

A new open-source project, Teleport-env, has been launched to provide an ultra-fast, OS-level snapshot and rollback solution for AI agents utilizing Monte Carlo Tree Search (MCTS) and reinforcement learning. Traditional methods, such as Docker containers, take several seconds to restart following an agent's destructive actions, which limits their utility for high-throughput AI experimentation. Teleport-env addresses this issue by enabling full-state recovery (filesystems and memory) in under 500 milliseconds, facilitating rapid testing and iteration of AI-generated scripts and commands. Teleport-env is built using a combination of overlayfs and Checkpoint/Restore In Userspace (CRIU), allowing for seamless environment restoration. The architecture employs a Cold Layer Switch to avoid slow file copying operations, providing a highly responsive testing environment. By leveraging Linux kernel capabilities, the system efficiently captures the state of a running agent, enabling it to return to a stable condition almost instantaneously after encountering errors. This reduced rollback time significantly enhances the performance of MCTS loops and reinforcement learning processes, making Teleport-env a powerful tool for developers working on autonomous coding agents.