Crab: eBPF-Based Checkpoint/Restore Runtime Achieves 100% Agent Recovery at 1.9% Overhead

arXiv·high signal

Crab bridges the agent-OS semantic gap for sandbox checkpoint/restore using an eBPF-based inspector that classifies OS effects per turn, a coordinator that aligns checkpoints with turn boundaries and overlaps I/O with LLM wait time, and a host-scoped engine for scheduling across co-located sandboxes. Results show recovery correctness jumps from 8% (chat-only) to 100%, checkpoint traffic drops up to 87%, and overhead stays within 1.9% of fault-free execution. The key insight: over 75% of agent turns produce no recovery-relevant state, enabling massive checkpoint sparsity.