Skip to content
AX Framework
Trust pattern

Sensitive Actions

Label actions by risk so agents and policies can treat them differently before anything happens.

AX proposalTrust & Control

Problem

Agents cannot tell a harmless action from a consequential one. Deleting an account and renaming a list look the same in an API reference.

Context

Use this pattern when:

  • Your product mixes routine and high-stakes actions.
  • Agents and their operators want to apply policies (approval, logging, limits) based on risk.

Recommended solution

Assign every action a risk level and machine-readable attributes: reversible or not, involves payment, shares personal data, affects other people. Publish these in the manifest so agents can plan, and enforce corresponding policies server-side.

  1. 01Define a small risk vocabulary (read, low, medium, high).
  2. 02Tag actions with attributes: reversible, financial, personalData.
  3. 03Publish attributes in the manifest.
  4. 04Enforce policies server-side regardless of agent behavior.

Interaction diagram

Interaction diagram for Sensitive Actions. 1. AI agent to Product · AX layer: GET /ax.json. 2. Product · AX layer to AI agent: delete_account: risk high, irreversible. 3. AI agent to AI agent: Policy: confirm with user first.AI agentProduct · AX layerGET /ax.json1delete_account: risk high, irreversible2Policy: confirm with user first
request response internal step

Example implementation

Illustrative example

Examples use a fictional store and illustrative field names. Adapt the shape to your stack; the qualities matter more than the exact format.

Risk attributesjson
{
  "name": "delete_account",
  "description": "Permanently delete the account and all associated data.",
  "risk": "high",
  "attributes": {
    "reversible": false,
    "financial": false,
    "personalData": "deletes",
    "affectsOthers": false
  },
  "requiresApproval": true,
  "agentAllowed": false
}

Accessibility and security

Accessibility

  • Use the same risk language in UI confirmation dialogs so the mental model is consistent for everyone.

Security

  • Some actions should not be available to agents at all; say so explicitly with agentAllowed: false.
  • Risk labels inform agents, but enforcement must not depend on agents honoring them.

Applies the principles Clear Permissions and Progressive Autonomy.