ARTFEED — Contemporary Art Intelligence

Crab: A Semantics-Aware Checkpoint/Restore Runtime for Agent Sandboxes

ai-technology · 2026-05-01

Crab is a transparent host-side runtime that bridges the semantic gap between autonomous agents and their sandboxed operating systems. Over 75% of agent turns produce no recovery-relevant state, making full per-turn checkpointing unnecessarily expensive. Crab uses eBPF to enable efficient checkpoint and restore without modifying agents or C/R backends.

Key facts

  • Crab is a semantics-aware checkpoint/restore runtime for agent sandboxes.
  • It addresses the agent-OS semantic gap.
  • Over 75% of agent turns produce no recovery-relevant state.
  • Full per-turn checkpointing is correct but too expensive under dense co-location.
  • Crab uses eBPF-based technology.
  • It does not require modifying agents or C/R backends.
  • The paper is available on arXiv with ID 2604.28138.
  • The work targets fault tolerance, spot execution, RL rollout branching, and safe rollback.

Entities

Institutions

  • arXiv

Sources