CORE GUIDE

MENTAL_MODELINTERMEDIATE6 min read

Reversible actions

Agent actions are safer when the system prefers operations that can be previewed, undone or compensated before granting irreversible authority.

Mental model

Reversibility is an action-design property: a proposed side effect should have a known rollback, compensation or staged-commit path whenever the business process allows it.

Why it matters

Permission design is not only about whether an agent may call a tool. Two allowed operations can have radically different blast radius: drafting an email versus sending it, preparing a refund versus settling money, or opening a pull request versus force-pushing main. Reversible intermediate states buy time for verification and human control.

01

Stage consequential effects before commitment

Classify tool actions by reversibility and consequence. Prefer read, preview, draft, soft-delete, pending or transaction-like operations where available. Put irreversible commit behind stronger evidence or approval, and define compensation for systems where true rollback is impossible. The agent runtime should know which transition crosses that boundary.

02

Example: delete becomes quarantine

Instead of granting an agent a permanent-delete tool, a document workflow exposes quarantine with a retention window. The agent can remove a bad document from active retrieval immediately, but operators can inspect provenance and restore it if the diagnosis was wrong. Permanent deletion is a separate audited transition.

Common failure modes

  • Giving one broad tool both preview and irreversible commit powers.
  • Calling an action reversible when rollback itself loses information or causes external effects.
  • Depending on a human to invent compensation only after an incident occurs.

Engineering heuristics

  • Prefer staged and compensatable operations for autonomous execution.
  • Make irreversible transitions explicit in tool contracts and UI.
  • Test rollback or compensation paths before increasing agent authority.

Takeaways

  1. 01Reversibility reduces the cost of a wrong decision.
  2. 02Permission scope should reflect consequence, not only API surface.
  3. 03Staged commitment creates room for verification and human review.

Reading evidence

UnseenPractice not completed

This records actions you actually took; it does not claim mastery, proficiency, or certification.

Used in

This Concept is reused across these canonical learning paths.

Related concepts from the Knowledge Graph

These relationships come from the canonical graph, not a separate Guide taxonomy.