Reproduce “DeltaBox: Scaling Stateful AI Agents with Millisecond-Level Sandbox Checkpoint/Rollback” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yunpeng Dong, Jingkai He, Shiqi Liu, Yuze Hou, Dong Du, Zhonghu Xu, Si Yu, Baochuan Yang and 2 more
DeltaBox provides coupled filesystem-and-process checkpointing for coding-agent sandboxes. It versions filesystem changes with overlay layers and resumes frozen process templates, coordinating both during rollback. This avoids repeatedly copying entire sandboxes. In a 30-iteration Qwen3-Coder-30B Django trajectory, DeltaBox took 1.01× the LLM-plus-action floor versus E2B’s 1.31×. This measures overhead, not coding success.
Key insights
- Treating consecutive agent states as marginally different allows for significant performance gains through change-based duplication instead of full-state copies.
- Coupling `fork()`-based process templating with dynamic, unmount-free `overlayfs` layer switching enables millisecond-level, atomic C/R for both process and filesystem states.
- Asynchronous memory page warming and proxying network connections are critical for preserving low-latency fast-path restores and template forkability in complex agent environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.