Training computer-using AI agents requires running many isolated sandboxes in parallel, each consuming gigabytes of memory. Slow tool executions in these sandboxes delay the entire training batch.
BunshinRL creates live copies of running workspaces that share common memory until a copy makes changes. It also caches and reuses results from repeated tool calls and predicts future calls to start them early.
It cuts peak memory by 43% and total memory by up to a half, while speeding up training by 23%. This makes training stateful agents with reinforcement learning much cheaper.