> ## 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.

# Create and configure an internal agent

> Create an internal agent, then set its instructions, guardrails, tools, Skills, model and subagents, and publish your changes.

This page walks through creating an internal agent and filling in each tab of its builder. Once it's configured, choose what starts it under **Runs on**. See [Internal agents](https://docs.conduit.ai/build-your-agent/agents/internal-agents/overview.md) for how the pieces fit together.

## Create the agent

**1. Open Agents**

Go to **Agents** in the workspace sidebar.

**2. Add an internal agent**

Select **Add agent > Create internal agent**. To describe the job in plain language and let Operator draft the agent for you, select **Build with Operator** instead.

**3. Name it**

Enter a short name that says what the agent does, such as "Arrivals checker", then select **Create**. Agent names must be unique in the workspace and can be up to 100 characters.

The new agent appears under **Internal agents** on the Agents page, with an **INTERNAL** badge in its sidebar.

## The agent sidebar

| Section | Item | Use it for |
| - | - | - |
| **Agent** | **Build** | The builder tabs: **Activity**, **Persona**, **Tools**, **Skills** and **Settings**. |
| **Runs on** | **Messages** | Channels the agent listens or replies on. See [Messages](https://docs.conduit.ai/build-your-agent/agents/internal-agents/messages.md). |
| | **Schedule** | Times the agent runs on its own. See [Schedules](https://docs.conduit.ai/build-your-agent/agents/internal-agents/schedules.md). |
| | **Webhook** | A private endpoint other systems call. See [Webhooks](https://docs.conduit.ai/build-your-agent/agents/internal-agents/webhooks.md). |
| | **Workflow** | Platform events the agent reacts to. See [Platform events](https://docs.conduit.ai/build-your-agent/agents/internal-agents/platform-events/overview.md). |

Each **Runs on** row shows its state at a glance, such as **Setup**, **Off**, **Active**, **2 active** or **3 channels**.

## Persona

The **Persona** tab holds what the agent is told on every run.

![An internal agent's header with its name and description, above the Persona tab showing Instructions in Formatted view](https://mintcdn.com/conduit-58d85182/Dxa79236R9TzscJw/images/internal-agents/persona-instructions.png?fit=max\&auto=format\&n=Dxa79236R9TzscJw\&q=85\&s=ab65aae70809fe0b9fc6519ac3cfaa4a)

Edit the agent's name and description at the top of the builder. Write the instructions in **Formatted** view, or switch to **Raw** to edit the markdown directly.

* **Instructions** describe the job: what the agent is for, what a good result looks like, who it reports to, and how to handle edge cases. Write them the way you'd brief a new teammate.
* **Guardrails** are rules the agent must always or never follow, such as "Never message a guest between 10 PM and 8 AM local time." See [Persona and guardrails](https://docs.conduit.ai/build-your-agent/agents/persona-and-guardrails.md).

> [!TIP]
>
> Keep instructions about the agent's role and goals. Put step-by-step procedures in [Skills](https://docs.conduit.ai/build-your-agent/skills.md), and facts in [Knowledge](https://docs.conduit.ai/build-your-agent/knowledge-base/overview.md). See [What goes where](https://docs.conduit.ai/build-your-agent/what-goes-where/overview.md).

## Tools

The **Tools** tab decides what the agent is allowed to do: connection tools from your integrations, custom tools, MCP tools, built-in Conduit tools, and AI-triggered workflows. Enable only what the job needs.

### Capabilities

The **Capabilities** list at the top of the **Tools** tab turns on general-purpose abilities that aren't tied to one integration. Which ones you see depends on your workspace.

![The Capabilities list with Browser, Firecrawl, Shell and Web fetch, each with an on/off switch](https://mintcdn.com/conduit-58d85182/Dxa79236R9TzscJw/images/internal-agents/tools-capabilities.png?fit=max\&auto=format\&n=Dxa79236R9TzscJw\&q=85\&s=92207cabf55e69d38ebb0dcc517ed639)

| Capability | What it lets the agent do |
| - | - |
| **Browser** | Navigate and interact with websites in the agent's browser. You can limit it to the domains you allow. |
| **Firecrawl** | Search, scrape and map websites through Conduit's Firecrawl connection. |
| **Shell** | Run commands in the agent's sandbox. |
| **Web fetch** | Read the content of a web page at a given URL. |

Turn a capability on with its switch.

An internal agent can only take actions its tools allow. If it needs to message a guest, create a task in your PMS, or post to Slack, enable that tool here. See [Tools](https://docs.conduit.ai/build-your-agent/tools.md) for every tool source and the built-in tools.

## Skills

The **Skills** tab assigns reusable procedures to the agent. A Skill tells the agent when a situation applies and which steps and tools to use, for example "How to rank today's cleanings". See [Skills](https://docs.conduit.ai/build-your-agent/skills.md).

![The Skills tab listing workspace Skills with a search box, an add button, and a switch on each Skill](https://mintcdn.com/conduit-58d85182/Dxa79236R9TzscJw/images/internal-agents/skills-tab.png?fit=max\&auto=format\&n=Dxa79236R9TzscJw\&q=85\&s=f06e541f997fda653306a5250b53f106)

Search the list and turn on the Skills this agent should use with their switches. Select **+** to add a Skill.

## Settings

Internal agents don't reply to contacts from an inbox, so their **Settings** tab has no **Escalation rules** or **Behavior** sections.

| Section | What it controls |
| - | - |
| **Details** | The agent's ID (with a copy button), who created it, and when it was last updated. |
| **Model** | The AI model the agent uses and its thinking effort. Not shown on every workspace. |
| **Subagents** | Which agents this agent can hand work to. See [Subagents](https://docs.conduit.ai/build-your-agent/agents/internal-agents/subagents.md). |
| **Widget overrides** | How the agent replies when a website widget triggers it. See [Agent settings](https://docs.conduit.ai/build-your-agent/agents/settings.md#widget-overrides). |

### Model

When your workspace shows the **Model** section, people who can manage agents choose the model from **Agent model**. Models are grouped by provider. For models that support it, pick a **Thinking effort**: **Minimal**, **Low**, **Medium**, **High** or **Extra high**. Higher effort lets the agent reason longer before acting, which suits complex, multi-step jobs.

The model you pick affects what each run costs. See [Credits](https://docs.conduit.ai/build-your-agent/agents/internal-agents/overview.md#credits).

## Publish your changes

Edits in the builder are staged. A save bar appears at the bottom of the page while you have unsaved changes. Select **Save** to publish them, or **Cancel** to discard them.

Select **History** in the top bar to review earlier published versions of the agent.

> [!NOTE]
>
> The **Activity** tab appears once the agent has been published.

## Turn the agent on or off

Open the status menu in the top bar. Under **Agent status**, pick **Active** ("Agent can run from enabled triggers") or **Inactive** ("Agent is turned off"), then select **Save**.

![The Agent status menu with Active selected and Inactive below it, above a Save button](https://mintcdn.com/conduit-58d85182/Dxa79236R9TzscJw/images/internal-agents/agent-status-menu.png?fit=max\&auto=format\&n=Dxa79236R9TzscJw\&q=85\&s=aaa2a55a75b448319fd964f8b750098a)

An inactive agent keeps its configuration and triggers. Its automatic triggers don't start it, but **Run now** can still start a real run with its configured tools.

## Test before you rely on it

Internal agents have no sandbox mode. Every run is real, and its tools take real actions. To test safely:

* Start with read-only tools, and add tools that change data once the agent behaves as you expect.
* Use **New chat** to give the agent a task and watch each step. See [Chat and activity](https://docs.conduit.ai/build-your-agent/agents/internal-agents/chat-and-activity.md).
* Use **Run now** to fire a mocked schedule, workflow or webhook trigger.
* Use **Try it now** in the schedule editor before you save a schedule, and pause a schedule with its switch while you adjust the agent.


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