> ## Documentation Index
> Fetch the complete documentation index at: https://docs.trellistech.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent permissions and approvals

> Check the sources behind an agent answer, and control which agent actions run immediately versus wait for a person

## What this helps you do

Trust an agent answer for the right reason — because you can see what it read —
and set each agent's boundaries so safe work runs at full speed while sensitive
actions wait for you.

## Read the sources behind an answer

Every substantive agent answer carries the material it worked from.

<Steps>
  <Step title="Open the sources">
    On a suggested reply, open **Sources** to see the records the agent used — property facts,
    documents, and conversation history each appear as their own source. A reply with no attached
    material shows "No sources are attached to this reply."
  </Step>

  <Step title="Check what each claim rests on">
    Match the answer's specific claims to the listed sources. **Sources used** names what actually
    informed the reply, not everything the agent could reach.
  </Step>

  <Step title="Inspect a full run">
    For agent work beyond a single reply, the run's **Evidence** shows what the agent did and found
    step by step.
  </Step>
</Steps>

If a claim has no matching source, or the source is a document you know is out
of date, correct the record first — see
[Fix wrong AI information](/guides/fix-wrong-ai-information) — rather than
editing the one answer.

## Control what runs immediately

An agent's boundaries live in its **Access & Safety** settings.

<Steps>
  <Step title="Set tool permissions">
    Every tool the agent can use is **Allowed**, **Ask first**, or **Off**. **Ask first** means the
    agent can draft the action, but a human must confirm before it runs.
  </Step>

  <Step title="Set work-area access">
    **View** and **Change** access are set separately per work area, so an agent can read one area
    while only proposing changes in another.
  </Step>

  <Step title="Keep deliveries reviewed">
    Outbound messages follow the reply flow's **Approve & send** control. For the situations that
    always deserve review, see [Approve a sensitive message](/guides/approve-sensitive-message).
  </Step>
</Steps>

## Where the line sits by default

The five action classes in
[Understand the five action classes](/guides/safe-operations-foundation)
explain the model behind these controls: reads and drafts are safe and run
immediately, while deliveries and spending pause for a person. Agent
permissions apply that model per agent and per tool — tighten a specific
agent's boundaries with **Ask first** or **Off** without slowing down every
other agent in the workspace.