> ## 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.

# How Conduit agents work

> What a Conduit agent does with an incoming message or call, the pieces you set up, and how you keep control of it.

Conduit handles customer conversations for you. It reads messages and calls coming in on every channel you use, answers the ones it can, takes action in your other systems when it needs to, and brings in a person when it should not handle something alone.

Every reply can be opened to show the reasoning and the sources behind it, so when something is wrong you can find the cause and fix it rather than guess.

## The three kinds of agent

| Agent | Who it talks to | What starts it |
| - | - | - |
| **Chat agent** | Your customers, over email, SMS, WhatsApp, OTA messaging, live chat and the other channels you connect | A new message arriving in the inbox it is attached to |
| **Voice agent** | Callers on a phone number you have set up for it | An incoming call to that number |
| **Internal agent** | Your team and your other systems, never customers directly | A teammate, a schedule, a webhook, or an event in your workspace such as a new reservation |

All three are built from the same pieces, described below. See [Agents](/build-your-agent/agents) for setup.

## The pieces you set up

| Piece | What it holds | Example |
| - | - | - |
| [**Knowledge Base**](/build-your-agent/knowledge-base/overview) | Facts the agent can look up | Check-in times, parking details, your cancellation policy |
| [**Skills**](/build-your-agent/skills) | Processes the agent follows when a situation comes up | How to handle a refund request or an early check-in |
| [**Tools**](/build-your-agent/tools) | Actions the agent can take in your other systems | Checking availability in your PMS, creating a cleaning task |
| [**Persona**](/build-your-agent/what-goes-where/style) | How the agent sounds | Warm and brief, signs off as the host team |
| [**Guardrails**](/build-your-agent/what-goes-where/guardrails) | Lines the agent must never cross | Never share the door code before a confirmed booking |
| **Behavior** | When the agent answers and when it hands over | Working hours, response delay, escalation rules |

Each piece has one job. Keeping facts in Knowledge, processes in Skills and tone in Persona means you always know where to look when you want to change something. See [What goes where](/build-your-agent/what-goes-where/overview) if you are unsure.

## What happens when a message arrives

<Steps>
  <Step title="The agent picks it up">
    A chat agent picks up a new message in its inbox. A voice agent picks up a call on its number. An internal agent starts when its trigger fires.
  </Step>

  <Step title="It looks up the relevant facts">
    The agent pulls in the Knowledge Base pages that relate to what is being asked. If the conversation is linked to a reservation, it uses the knowledge for that listing and its groups as well as anything that applies workspace wide.
  </Step>

  <Step title="It checks whether a Skill applies">
    If the situation matches one of the agent's Skills, such as a late checkout request, the agent follows the steps in that Skill.
  </Step>

  <Step title="It uses a tool if it needs to">
    If the Skill calls for an action or live information, the agent uses one of the tools it has been given, for example checking availability or creating a task.
  </Step>

  <Step title="It checks the reply">
    The draft is checked against the Guardrails and written in the voice set by the Persona.
  </Step>

  <Step title="It sends, drafts, or escalates">
    On autopilot the reply is sent. On copilot it waits as a draft for your team to approve. If an escalation rule applies, or the agent is missing information, the conversation goes to a person instead.
  </Step>
</Steps>

## How you stay in control

You can open any AI reply in the Inbox to see which Knowledge pages it used, which Skill it followed and which tools it called. If the answer was wrong, you can fix the source directly from there so the agent does better next time.

Changes to an agent are staged and only go live when you publish them, and you can try them in the Sandbox first. Most teams start on copilot, where a person approves every reply, and move to autopilot one topic at a time. See [Road to Autopilot](/build-your-agent/road-to-autopilot/overview).

## Around the agents

**Workflows** run automations that do not need judgment, such as sending check-in instructions the day before arrival. See [Workflows](/workflows/getting-started).

**Operator** is an assistant for your team. It helps you build agents, write Skills, find gaps in your Knowledge Base and pull reports. See [Operator](/build-your-agent/operator).

**Insights** show your automation rate, why conversations escalate, and the conversations behind every number. See [AI Automation Dashboard](/insights/ai-automation-dashboard).

<CardGroup cols={2}>
  <Card title="Agents" icon="sparkles" href="/build-your-agent/agents">
    Create chat, voice and internal agents.
  </Card>

  <Card title="Knowledge Base" icon="book-open" href="/build-your-agent/knowledge-base/overview">
    Manage the facts your agents look up.
  </Card>

  <Card title="Skills" icon="cube" href="/build-your-agent/skills">
    Write the processes your agents follow.
  </Card>

  <Card title="Tools" icon="wrench" href="/build-your-agent/tools">
    Control what actions your agents can take.
  </Card>
</CardGroup>


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