Skip to main content

​
What this helps you do

Predict which actions in Trellis run immediately and which pause for a person, so you can set up permissions and agent access with confidence. Every action a teammate or AI agent takes belongs to one of five classes, and the controls in Settings > Team & Permissions and each agent’s Access & Safety follow the same model.

​
The five classes

A wrong change can be corrected inside Trellis. A wrong delivery has already reached a guest or an external system. That is why reads, drafts, and changes run at full speed while deliveries and spending pause for review.

​
Where the controls live

1

Roles

Go to Settings > Team & Permissions. Each role is a matrix of resources and actions — View, Create, Edit, Delete, Assign, and Approve. Property access notes: “Restrict which properties this role can see and act on. Empty selection = all properties.”
2

Agent tool access

In an agent’s Access & Safety, every tool is Allowed, Ask first, or Off, with View and Change access set separately per work area.
3

Agent autonomy

Above tool access sits the autonomy mode: Observe only, Ask before acting, Act within limits, or Paused. Even in Act within limits, tools set to Ask first still stop for a person.
4

Deliveries

Drafted guest replies wait as suggested replies for Approve & send. Agent actions set to Ask first stop at a Requires approval prompt showing exactly what would run.

​
Judge each request on its class

Before an unfamiliar action, ask which class it is. If the answer is deliver or spend, expect an approval — and treat declining as normal operation, not a failure. The prompt shows the exact content and input, so judge each item on its own evidence.

​
Common problems

Review the tool’s setting in Access & Safety. Tools set to Allowed run without a prompt in Act within limits; move sensitive tools to Ask first.
Their role grants View but not Edit on that resource. Review the role matrix in Settings > Team & Permissions.
Reads, drafts, and changes never require approval. If routine internal edits are prompting, the affected tools or permissions are set more strictly than the action class requires.