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

# Agent settings

> Escalation rules, language, working hours, reply preferences, and widget options for an agent.

Open an agent and select the **Settings** tab to control how it operates once it knows what to say.

| Section | What it controls |
| - | - |
| **Details** | The agent's ID, who created it, and when it was last updated. |
| **Subagents** | Which internal agents this agent can hand work to. |
| **Escalation rules** | What happens when the agent escalates a message. |
| **Behavior** | Language, voice, working hours, reply preferences, and signature. |
| **Widget overrides** | How the agent replies in the chat widget. |

Internal agents don't reply to contacts, so they don't show **Escalation rules** or **Behavior**.

## Subagents

Choose which internal agents this agent can delegate work to. The agent hands a task to a subagent when it needs that agent's specialized tools or knowledge. Only internal agents appear in the list.

## Escalation rules

Each rule says **When** a message matches its conditions, **Then** take an action. Conditions are escalation types, such as **AI missing info**, **AI needs your help**, or **Conflicting info**, and conversation tags.

| Action | What happens |
| - | - |
| **Escalate** | The conversation goes to a team member and the agent doesn't respond. |
| **Answer and escalate** | The agent responds and asks for the information it couldn't find. |

Select **Add rule** to create one, or hover over a rule and select edit to change it.

## Behavior

### Language and voice

| Setting | What it does |
| - | - |
| **Language** | Replies in the original text, matches the contact's language, or translates to a language you pick. |
| **Voice** | Lets guests talk to the agent in the chat widget. Pick a voice, or leave it **Off**. |

### Working hours

Select **Agent working hours** to set when the agent runs.

### Reply preferences

| Setting | What it does |
| - | - |
| **Delay** | Waits before the agent's reply is sent, so your team has time to step in. 10 seconds is recommended. |
| **Auto Mark-as-Done** | Lets the agent close conversations it has finished. |
| **Auto Mark-as-Done Instructions** | Defines what counts as still owed before a conversation can close. |

#### Auto mark-as-done

After the agent handles a message, whether it replies or stays silent, it marks the conversation done only when nothing is still owed to the contact. Escalations always stay open.

![Reply preferences with Auto Mark-as-Done turned on and Auto Mark-as-Done Instructions using Conduit's default](https://mintcdn.com/conduit-58d85182/-4eorIAjTHfeFOCj/images/auto-mark-as-done-settings.png?fit=max\&auto=format\&n=-4eorIAjTHfeFOCj\&q=85\&s=aa91fd90efcebff1a350dc2263a4f455)

When Auto Mark-as-Done is on, **Auto Mark-as-Done Instructions** define what counts as still owed. By default, **Use Conduit's default** is on and the agent keeps a conversation open for:

* An unanswered escalation.
* An action you or your team promised that has not happened yet, even one the contact already accepted.
* Information needed from the contact.
* A request a rule or guardrail told the agent to leave for your team, which nobody has answered yet.

An upcoming stay, a possible future message, and sign-offs like "let us know if that helps" do not keep a conversation open.

To write your own rules, turn off **Use Conduit's default**. The editor starts from Conduit's default text so you can adjust it. Turn the switch back on to return to the default.

> [!TIP]
>
> Keep custom instructions focused on what your team still owes the contact, for example "A refund we said we would process."

### Signature

Text added to the end of every reply this agent sends.

## Widget overrides

These settings apply only when the agent replies in the chat widget.

| Setting | What it does |
| - | - |
| **Stream responses** | Shows the reply word by word as the agent writes it, instead of as one finished message. |
| **Skip conversation tag classification** | Replies without tagging the conversation first, since a person is waiting. Tagging never runs while streaming. |
| **Skip structured knowledge preload** | Doesn't load structured knowledge up front. The agent still looks it up when it needs it. Preload never runs while streaming. |

Changes on this tab are staged. Review them in the save bar and publish when you're ready. **Voice** is the exception: it saves as soon as you pick one.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.