Definition
Plain language
Saving the full state of an AI agent's environment so it can rewind if it makes a mistake.
As stated in the literature
The operation of capturing combined filesystem and process-memory state of a running sandbox atomically, supporting later rollback to that exact state for tree-search agents and RL training.
Also called: sandbox checkpointing
Why it matters: Cheap state snapshots are what make tree-search and reproducible RL training possible for agents that interact with full computer environments.
For example, before letting the agent run a risky `rm -rf` command, the harness snapshots both the filesystem and the running processes so it can rewind if things go wrong.
Heard on the show
“They get sandbox checkpoint down to about fifteen milliseconds.”Episode 068 — The OS Trick That Makes Tree Search Practical for Coding Agents