Can You Undo That? Reversibility Grading of AI Agent Operations
Can You Undo It? Reversibility Classes for AI Agent Actions
The authors propose Reversibility Classes, dividing tool-using agent operations into five levels by degree of reversibility, with accompanying parameter-sensitive rules, conservative defaults and prior-state capture conventions; a version-locked implementation and evidence-linked registry cover 536 tools across 20 MCP servers, and all 738 rule test cases met their declared expectations. Across 2,158 recorded runs, session guardrails blocked the relevant operation in all 29 attributable attack-positive runs, and the read-write proxy in 18, while adding 9 approval points across 1,209 benign runs. Controlled experiments on files, Git, SQLite and three browsers show that recovery depends on captured state, subsequent changes, editor storage and the recovery path, and expose one Git merge-state exception. The authors state that the evidence supports making recovery assumptions explicit and testable, but cannot prove the registry is accurate as a whole or that it reduces harm in production environments.