← News

Agents & MCPSources: arXiv · 16 Nov 2026

DeltaBox brings millisecond checkpoint and rollback to AI agent sandboxes

A research paper describes an OS-level design that checkpoints and restores agent sandboxes in milliseconds by copying only state changes instead of full snapshots.

In plain words

AI agents that explore many possible actions need to save and restore their entire working environment — files and running processes — very quickly. Existing tools copy everything each time, which takes hundreds of milliseconds or more. DeltaBox instead records only what changed between saves, cutting checkpoints to about 11 milliseconds and rollbacks to under 2 milliseconds in the paper's tests.

What was announced

The arXiv paper, by the DeltaBox authors, argues that agent workloads such as test-time tree search and reinforcement learning depend on frequent state exploration, which requires rapid checkpoint/restore of the complete sandbox, including files and process state. Existing mechanisms duplicate the entire state and take hundreds of milliseconds to seconds per cycle, which the authors identify as a bottleneck for deep search and large-scale fan-outs. Their key observation is that consecutive sandbox states differ only marginally, but the lack of operating system support has made change-based copying hard to realize.

The proposed abstraction, DeltaState, combines two co-designed OS mechanisms. DeltaFS checkpoints file state by freezing the writable layer and inserting a new one, so updates become copy-on-write and rollback becomes a layer switch. DeltaCR restores process state by forking from a frozen template process, and a coordinator pairs the two so each checkpoint captures one consistent filesystem-and-memory state. On SWE-bench, the authors report DeltaBox hides its 10.83 ms checkpoint work under inference and rolls back in a mean of 1.86 ms, with the same warm-template fork supporting low-cost RL rollout fan-out. The system is open-sourced on GitHub.

What it does not show

  • The evidence is the paper's abstract only; the reported timings are the authors' own benchmarks on SWE-bench and have not been independently confirmed.

For security and IT teams

Scope
The paper covers sandbox checkpoint/restore for LLM agent workloads such as tree search and reinforcement learning, not general-purpose backup tooling.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.