Skip to Content
AI AgentPending Actions

Pending Actions

When the agent wants to make changes, it creates a pending action that requires your approval. This ensures the agent never takes destructive or unintended actions without human oversight.

How It Works

  1. You ask the agent to do something (e.g., “tag all pending orders as priority”)
  2. The agent creates a pending action with details about what it will do
  3. You review the action — including the scope and affected items
  4. You approve or reject the action
  5. If approved, the action executes

Pending Action Details

Each pending action includes:

FieldDescription
scopeWhat will be affected (e.g., “42 orders”)
actionWhat will happen (e.g., “add tag: priority”)
affectedItemCountNumber of records that will change
idempotencyKeyEnsures the action can only execute once
expiresAtActions expire if not approved within a time window

Approving Actions

Portal

Click Approve or Reject on the pending action card in the chat.

API

POST /api/agent/actions/:id/approve

or

POST /api/agent/actions/:id/reject

Safety Guarantees

  • Actions with side effects always require approval
  • Read-only queries execute immediately (no approval needed)
  • Each action has an idempotency key — approving twice won’t double-execute
  • Actions expire after a configurable time window
  • The agent shows exactly what will happen before you approve