Making AI agents 100× faster by saving only what changed

AI agents doing tree search or reinforcement learning need to constantly save and restore their entire state—files, memory, process context. Existing approaches copy everything, costing hundreds of milliseconds per operation. DeltaBox observes that consecutive checkpoints differ only slightly, so it saves only the deltas instead. Two new OS mechanisms—DeltaFS for filesystem layers and DeltaCR for process snapshots—reduce checkpoint time to 14ms and rollback to 5ms. On SWE-bench and RL tasks, this 50–100× speedup lets agents explore orders of magnitude more solution branches under the same time budget.