Skip to main content
Choose what an agent can read, change, and send, and which actions need approval. Use sources and AI Trace to inspect an answer or action.

​
Read the sources behind an answer

Check sources on a sent message in Trellis
Open Sources or a citation on the reply when available. Read the property record, document, or other material behind a specific claim. Sources used refers to the material used for that answer, not everything the agent could access. Some replies have no attached sources.
Inspect an AI Trace in Trellis
Open AI Trace or the run’s Evidence for more detail about retrieved information, tool calls, and results. If the source is missing or wrong, follow Fix wrong AI information.

​
Control what runs immediately

Open the agent’s Access & Safety settings. Review its work-area View and Change access and each tool’s setting:
  • Allowed: available to run within the agent’s other permissions and autonomy limits.
  • Ask first: prepare the action and wait for a person’s confirmation.
  • Off: unavailable to that agent.
Review the autonomy mode as well. Observe only, Ask before acting, Act within limits, and Paused set the agent’s broader behavior. A tool set to Ask first remains reviewed even when the agent can act within limits elsewhere. When an approval appears, inspect the exact record, recipient, content, and change. Approve only that intended action; decline or clarify when the target or result is wrong.

​
Autopilot and workflow approvals

The five action classes help identify actions that need review. Configure approvals explicitly; the category alone does not enforce them. With Autopilot off, suggested guest replies remain for your review and Approve & send. Workflows have their own approval and timeout paths; follow Approve a sensitive message when a message must never send without a decision.

​
Check the boundary

Test a read, a draft, and an action that should require approval using a controlled record. Confirm the allowed work completes and the reviewed action pauses before executing. Open the resulting record or message after approval to verify the outcome.