Risk Levels

Agent Receipts classify every action into one of four risk levels. Risk levels are used for filtering, alerting, and authorization policy.

The four levels

Level	Description	Examples
low	Read-only or easily reversible	Read a file, navigate to a URL, create a draft
medium	Modifies state but reversible or low-impact	Edit a document, move a file, modify settings
high	Significant state change, may be hard to reverse	Send an email, delete a file, share a document
critical	Financial commitment or irreversible action	Make a purchase, authorize a payment, delete an account

Default assignment

Risk levels are assigned by action type as defaults in the Action Taxonomy. Each action type has a default risk level that reflects the typical severity of that action.

Escalation rules

Implementations may escalate but must not downgrade risk levels based on runtime context.

For example, a filesystem.file.delete action has a default risk of high. If the file being deleted is a system backup, the implementation should escalate the risk to critical. But an implementation must never downgrade a high-risk action to medium, regardless of context.

The no-downgrade rule ensures that risk levels serve as a reliable floor for compliance and audit tooling. A consumer of receipts can trust that any receipt marked high represents at least a high-risk action by the taxonomy’s definition.

Use cases

• Filtering: Show only high and critical actions in an audit dashboard
• Alerting: Trigger real-time notifications for critical actions
• Authorization gates: Require explicit user confirmation before executing high or critical actions
• Compliance reporting: Generate reports filtered by risk level for regulatory requirements