Action tiers
Everything the agent can do falls into one of two tiers:- Auto — safe, reversible, or read-only actions (looking up a delivery, summarizing performance, drafting a message) run automatically.
- Confirm — destructive or customer-facing actions (canceling, refunding, reassigning a provider, rescheduling, sending notifications) surface a confirmation card first. Nothing happens until you approve it.
Confirmation always wins. Even when an agent runs in auto-execute mode (see Scheduling & execution and the Entity monitor), actions in the Confirm tier still pause for human approval before anything is applied.
The confirmation flow
When the agent proposes a Confirm-tier action, it shows a card describing exactly what it will do — the entity affected, the action, and the parameters. You can approve it to run, or decline it. Approving runs the action with the same audited permissions you have in the Portal.Per-organization overrides
The defaults are sensible, but admins can tune which actions sit in Auto vs Confirm for their organization — tightening guardrails for sensitive operations, or loosening them for trusted, high-volume workflows. One exception: reaching a customer or driver by text or voice call always requires confirmation and can’t be moved to Auto. Individual custom agents can be tightened further: an agent’s confirm-first actions list forces confirmation for specific actions on top of whatever your organization already requires — tightening, never loosening, the org default. See the Custom agent configuration reference for the recognized action keys.Related
Scheduling & execution
Run agents on a schedule or trigger in report-only or auto-execute mode.
Entity monitor
Watch deliveries and act on changes — within your guardrails.
Custom agents
Define reusable, scoped agents — and force confirmation for specific actions per agent.
Channels: Slack, SMS & voice
Work with Nash from Slack, and reach customers or drivers by text or voice — always with your approval.