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

# Agents

> Let Marro's AI team plan next steps, analyse calls and keep conversations going.

Agents work on your leads in the background, without anyone asking: they reply, call, follow up and update leads on their own. You set what each one is allowed to do, every action is recorded on the lead, and your team can take over a chat or change what an agent does at any time.

> [!NOTE]
>
> Agents are part of Marro AI, on Growth and Complete. Changing their settings needs permission to manage settings.

## Your AI team

No agent is on by default. Your AI team is the agents your business adds. Open **AI → Agents** (its tabs are **Agents**, **Knowledge** and **Settings**) to see your own agents, one card each: its colour, whether it is **Working** or **Off**, what it does, its **runs this month** and when it **last ran**, with **Activity** and **Edit**. **Add an agent** is at the right of the tab row. With more than three agents, a search box and a status filter appear.

Choose **Add an agent** and describe the job. Marro proposes an agent for you. **Templates or manual setup** is available if you prefer to start from a template or **Start from scratch**:

| Template | What it does |
| - | - |
| **First caller** | Calls every new lead within minutes with AI. If they don't answer, it sends a WhatsApp and books a follow-up for the owner. |
| **Chat receptionist** | Answers leads on WhatsApp, Instagram, Messenger and the website chat, day and night, and hands over to a person when it should. It works with any one of them connected; you don't need a WhatsApp number. |
| **Email assistant** | Reads new lead emails and drafts threaded replies for review. Routine replies may be automatic only after you explicitly enable Email → On its own. |
| **Follow-up coordinator** | Sets the next step after every call, chat and note. |
| **Document collector** | When a lead reaches the stage you choose, asks for their documents through the portal and chases until they arrive. |
| **Call reviewer** | Reads every recorded call, AI calls and your team's calls, and saves what was said. |

Picking a template fills in the form. Nothing is on until someone chooses **Switch on**. Industry suggestions are labels only. Each card shows what the agent did this month. A template that isn't in your plan, or needs something set up first (AI calls, a connected chat, the client portal), says so. You can also ask the [assistant](https://docs.marro.si/guides/assistant.md) to switch an agent on or off.

When your organization first got agents, the switches you already had became agents automatically. The call reviewer stays on only for businesses that had calls analysed in the last 30 days.

## Create an agent

### Describe the job

Choose **Add an agent**. In **What should your agent do?**, write the result you want, for example: “When a new lead comes in, understand what they need and prepare a WhatsApp welcome message.” You can pick a starting example instead of writing from scratch.

Choose **Plan my agent**. Marro suggests a name, job instructions, when it starts, and the abilities it needs. **Chosen for this job** explains each ability and whether it acts **On its own** or **asks you first**. It checks your plan, permissions, connected services and which abilities that kind of agent actually supports. Customer-facing and money abilities start with **Ask me first**. It does not connect services or grant permissions for you.

If something important is unclear, **A little more detail** asks up to three questions. Answer in **Answer or change something**, then choose **Update the plan**. Use the same field to refine a proposal, for example “Create a task for the owner instead of sending a message.” Apply or clear any typed changes before choosing **Review settings**.

**Before you start** lists setup gaps and limitations. If you already have a chat receptionist, follow-up coordinator or call reviewer, the plan points you to **Edit** that agent. Planning cannot create a second one of those kinds. Chat receptionists answer connected chats and can hand over; their optional actions are currently notes, follow-ups and lead details. Direct appointment booking, quotes and calls are not extra chat-receptionist abilities yet. Email assistants use the separate **A lead emails a connected mailbox** trigger and support Email, Notes and Follow-ups only; they never reply through social chats. See [Email replies](https://docs.marro.si/guides/email.md#let-ai-help-with-replies) for sender access, review and exclusions.

Choose **Review settings** once required questions and blockers are resolved. The next screen summarises the draft; open **Customise instructions and abilities** to change details or approval choices. **Save, off** saves it without starting work. **Switch on** saves and activates it. Planning alone never saves or activates an agent and never contacts a customer; planning and trials use your AI allowance. If planning is unavailable, your description stays in the form and you can retry or choose **Templates or manual setup**.

### Review the settings

Whether Marro planned the agent or you start from a template or from scratch, its settings have these sections:

1. **Its job:** brief it in plain words, as you would a new team member. **Write it with Marro** or **Improve it with Marro** helps with the wording.
2. **When it works** (the trigger): a new lead, a lead messages in a chat, after a call, chat or note, a call has a recording, a stage, a document upload, a booking or a payment, at once or after a wait.
3. **What it may do:** choose what the agent is allowed to do. A line at the top says in words what it can do. Abilities are in groups (the lead and the team, messages to the lead, calls and documents, appointments, quotes and payments, email sequences, reports) with a search box; each has a one-line description and says how far it reaches: **Looks things up**, **Changes your records**, **Reaches your customers** or **Involves money**. Tick an ability, then choose **Ask me first** or **On its own**. Anything that **Involves money** always asks you first. These permissions are the agent's own; other agents don't get them. Past eight abilities the form suggests splitting the agent in two, because focused agents work best. You can leave them all off for a **Call reviewer** and still save its instructions.
4. **Most actions a day** (optional): the most the agent may do in a day. Leave it empty for the usual limit (50, of which 20 may reach customers). When it reaches the limit it stops for the day and its **Activity** says so.

### Try its next actions

For an agent that works on leads or follows up after calls, chats and notes, **Try its next actions** lets you choose a recent lead and select **Try on this lead**. The trial proposes its next actions without changing customer records or sending messages. It does not simulate trigger timing or delivery. A chat receptionist or call reviewer does not use this trial: it does not simulate a live chat, voice call or call review. Changing the instructions, abilities or chosen lead clears the displayed trial result.

An agent never does more than the person who owns it could: it works with that person's permissions and your plan, and an ability they can't use isn't offered. The owner is the person who created the agent; saving an agent whose owner has left makes you its owner.

Your business can have one agent for each of these three: a lead messages in a chat, after a call, chat or note, and a call has a recording. If another agent already does the same step on the same trigger, the form says so. Agents never repeat each other: when one agent has already done the same kind of thing to a lead in the last 12 hours (a WhatsApp, an email, a call), the next one's attempt waits for a person.

Each agent runs as an automation of its own, so its runs show in Automations too. Each agent has its own colour. Press the coloured dot on its card to choose another from twelve; administrators only. The agents take your company facts from [business knowledge](https://docs.marro.si/guides/business-knowledge.md).

Choose **Edit** on an agent to change its instructions, trigger, permissions and extras; the switch on its card turns it on or off, and its menu has **Delete**. Some agents have extras there:

* **Follow-up coordinator**: **Plan next steps**, **Draft WhatsApp replies** and a **Daily briefing** for each person.
* **Chat receptionist**: **Replies per contact** caps its replies to one person in a rolling 24 hours.
* **Chat receptionist**: **Follow-ups** lets it check back on a chat that is waiting on the person (they said they would think about it, ask their parents or send documents). It is off until you turn it on. It sets at most one check-back per chat, between 1 and 72 hours ahead, follows up at most twice a week, and drops the check-back as soon as the person writes. It never follows up after "stop" or "not interested", only inside WhatsApp's 24-hour window, and it can decide to send nothing.

## Give them instructions

Each agent has instructions, under **Manage**: how it should work for your business, such as its tone, what to ask and what to listen for. When your organization has AI calling, the **Voice agent** is listed too.

* Leave them empty and the agent follows the **Best practices** listed below the field.
* Press **Add** on any best practice to copy it into your own, or **Add all to edit** to start from all of them.
* Your instructions replace the best practices. **Clear** goes back to them.

Some rules always apply, whatever the instructions say: facts come only from your business knowledge, customers' messages are never taken as instructions, and the agents hand over to a person when they should. Up to 4,000 characters per agent.

You can also ask the [assistant](https://docs.marro.si/guides/assistant.md) to change an agent's instructions; it shows you the new text before saving it.

## Choose what they may do

In an agent's **Edit**, tick the abilities it may use and set each to **Ask me first** or **On its own**. Each ability can be:

* **Off** (not ticked): the agent never does it.
* **Ask me first**: the agent prepares it for a person to check first.
* **On its own**: the agent acts straight away.

How far an ability reaches decides what you may choose:

* **Looks things up** and **Changes your records** (notes, follow-ups, lead details, tags, reading a quote, a report or an appointment's availability) may be **On its own**.
* **Reaches your customers** (a message, a call, a stage or owner change, booking or rescheduling an appointment) may be **On its own** once you choose it. Until then it waits for a person.
* **Involves money** (a quote, an invoice, a payment link or reminder) always **asks you first**.

An ability that can't be used yet is hidden, or says what to set up first (connect WhatsApp, switch on AI calls, turn on the client portal); one outside your plan isn't offered. You can also describe an agent to the [assistant](https://docs.marro.si/guides/assistant.md), in the app or staff WhatsApp. It uses the same planning checks, explains the proposal and asks for approval before creating or changing the agent. Creation defaults to saved off unless you ask to switch it on; the approval names the trigger, wait, daily limit, abilities and instructions.

## See their work

Choose **Activity** on an agent for its log: newest first, grouped by day, each entry saying in plain words what it did, for which lead, what changed and how it went (**Done**, **Waiting for you**, **Skipped** or **Did not finish**). Filter by **All**, **Waiting for you** or **Did not finish**; **How it works** shows what starts it, what it may do and when it hands over. Everything an agent does is also recorded on the lead.

Every suggestion waiting for review is listed under **Waiting for review** at the top of **AI → Agents** (suggestion, lead, why, when), newest first: **Approve** (the button says what it does, such as Send) or **Dismiss**. An approval carries out exactly what was suggested, once.

## Settings for all agents

**AI → Settings** holds the [assistant's WhatsApp settings](https://docs.marro.si/guides/assistant.md#for-administrators), the **Agent decides** modes for [automations](https://docs.marro.si/guides/automations.md), and **Pause agents**.

## Pause them

**Pause agents**, on **AI → Settings**, turns off all your agents and stops their queued work. Actions already started may finish. The assistant, live calls and automations keep their own settings.

## Hand a single chat to the AI

In the [Inbox](https://docs.marro.si/guides/inbox.md), **Let the AI reply** gives one chat to the conversation agent, and **Take over** takes it back.

## Lead memory and current facts

A lead’s case file preserves continuity between runs. Newly saved facts include a timeline activity, the exact supporting words and observation date. Marro checks that the activity belongs to that lead and that the quote contains the recorded value. Facts expire after at most 90 days; expired or unsupported facts are excluded when reading the file. Current lead fields and operational reads take priority over historical summaries.

Customer-agent retrieval excludes staff-only, expired, future-dated and superseded knowledge. Listed prices come from the current visible product catalogue. Historical summaries cannot establish a current payment status, price or appointment time. If a required operational read is unavailable, the agent cannot confirm that specific fact and can offer the existing callback or booking flow.


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