Can You Undo It? Reversibility Classes for AI Agent Actions
Summary
- Researchers introduced Reversibility Classes, a declared five-level vocabulary with argument-sensitive rules and prior-state capture contracts for tool-using AI agents.
- The implementation covers 536 tools across 20 MCP servers, with all 738 rule-case checks conforming to declared expectations.
- Session guard experiments gated actions in 29 attack-positive runs while adding nine approval points across 1,209 benign runs.
- Controlled experiments across files, Git, SQLite, and browsers show recovery depends on state capture, subsequent changes, and recovery routes, exposing a Git merge-state exception.
Why it matters
Understanding and classifying whether an AI agent's tool actions can be undone is critical for preventing unrecoverable errors and managing risk in autonomous execution environments.