> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mindosoftware.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Create an AI agent — Complete guide
> How to build an AI agent in Mindo end to end - the prompt, your business policies, capabilities, testing, settings and change history
An **AI agent** is the assistant that automatically replies to your customers on WhatsApp, Instagram or the web widget. In this guide you'll build one from scratch: create it, write its prompt, load your business rules, add capabilities, test it and get it ready for production.
Everything happens under **Messaging → AI Agents** (`/dashboard/ai-agents-management`).
## Demo video
The video is narrated in Spanish, but the on-screen steps match this guide exactly.
## Before you start
* You need access to the **AI Agents** section of your Mindo account.
* For the agent to actually reply you'll need a **connected WhatsApp session**. You can create the agent first and connect a session later.
## Create the agent
In the left menu, under **Messaging**, click **AI Agents**. You'll see every agent in your account with its status (Active / Inactive) and three shortcuts per card: **Test agent**, **Quick Chat** and **Settings**.
Click **Create New Agent**. Mindo asks which type you want:
* **Ecommerce** (recommended): asks you a few questions about your business and **auto-generates the prompt** by analysing your website. Ideal for online stores.
* **Custom**: starts blank and you configure everything yourself.
**ECommerce Custom**, **Flow** and **Import agent** are for advanced cases. If you need them, talk to your Mindo contact.
Type the **Agent Name** (at least 3 characters) and click **Create New Agent**.
The agent now exists. At the top you have the **status**, a **search box** (`⌘K`) that finds anything inside the agent, and the **Deactivate**, **Publish** and **Full history** buttons. On the left, the six tabs: **Prompt**, **Knowledge**, **Capabilities**, **Testing**, **Settings** and **History**.
## Prompt
The **prompt** defines who the agent is and how it speaks.
On the **Prompt** tab, click **Edit** on the *System Prompt* card and describe in plain language who the agent is, how it should talk and what it can resolve. It's Markdown, so you can use lists and bold.
Click **Save**. Notice the selector that reads **Development**: you're editing the test version, so nothing reaches your customers yet. It only goes live when you **publish**.
The **History** button stores every version with its date and author. If a change didn't work out, you can go back to the previous version without losing anything.
Further down, **Task Flows** makes the agent follow a fixed sequence of steps — for example, always asking for the order number before answering.
## Knowledge: structured policies
**Structured Policies** are your concrete business rules, written clearly so the agent can use them when replying. Keeping them separate from the prompt means you can update a deadline or a condition without touching the agent's personality.
Click **Add policy**. A form opens with four fields:
| Field | What to put |
| --------------------------- | ----------------------------------------------- |
| **Title** | What the policy is about |
| **Type** | The category to group it under (e.g. *Returns*) |
| **Description** | The full rule, with the exact details |
| **Usage Context / Trigger** | When the agent should apply it |
The policy is stored, grouped by type and with its own toggle to enable or disable it. You can edit or delete it anytime.
Above the list you have the **environment selector**, the policies' own **Publish** button, and **Import / Export**. If your rules are already in a spreadsheet, import them (JSON or CSV) instead of typing them one by one.
## Capabilities
This is where you give the agent concrete powers.
* **Tools**: actions the agent can run, like checking stock or an order status. Click **Add new tool** to browse the catalogue.
* **Guardrail**: enabled by default. It classifies greetings and goodbyes so the agent doesn't over-reply. Turn it off and the agent always answers.
* **Follow-ups**: automatic messages to re-engage conversations that went quiet.
How to configure follow-ups so the agent messages back customers who never replied.
## Testing
* **AI Simulation**: create test customers with different personalities. You can **auto-generate them from the prompt** or **create them manually**.
* **Standard Tests**: load fixed questions with the answer you expect, click **Run All** and Mindo tells you which ones passed.
* **Reviews**: **Create Review** generates a review so your team can rate the agent's interactions.
### Quick Chat
The fastest test: you write as if you were a customer and see how it answers. Under each reply you get 👍 / 👎.
Thumbs down asks you for a reason and **creates a ticket** for the Mindo team: it does not change the agent instantly. If you want an immediate change, edit the prompt or the policy yourself.
Test the agent and leave feedback step by step.
How to track your corrections.
Rate the agent's interactions.
## Settings
### Sessions and Audience
The **session** is the line the agent replies through. Click the **+**, tick the line and connect it. The **Audience** is optional and limits which contacts the agent responds to.
Connect an API number.
Link your Instagram account.
### Agent settings
One card per setting:
| Setting | What it's for |
| ----------------------------- | ---------------------------------------------------------------------------------------- |
| **Agent Schedules** | When the agent is available (see [Branches and holidays](/en/guias/sucursales-feriados)) |
| **Response Settings** | Format and style of the replies |
| **Grammar Rules** | Grammar and spelling rules |
| **Conversation Memory** | How much context the agent remembers |
| **Document Extraction (OCR)** | Automatic reading of documents the customer sends |
| **AI Handoff** | Which human it escalates to, by reason, audience and channel |
| **Interview Answers** | View and manage what the agent collected |
**Model Settings** and **Endpoint Settings** are advanced: the Mindo team usually sets them up for you.
### Clone, export or delete
At the bottom of the tab you have **Clone Agent** (identical copy), **Export Agent** (JSON backup) and **Delete Agent**.
**Deleting an agent is permanent.** There's no undo — export it first if you want to keep the configuration.
## History
Every change is recorded: who did it, when and what changed. Filter by **Settings**, **Flows**, **Tools**, **Policies** or **Script**, and the **Live** indicator shows changes as they happen.
The **Full history** button in the header opens the complete log with every change together. It's the best place to understand why the agent started answering differently.
## Publish and turn it on
Click **Publish**. Mindo shows you the **Development → Production** diff so you can review it before confirming.
The **Publish** button is disabled if you have **unsaved changes** or if the agent is **inactive**.
With the status set to **Active** and at least one **connected session**, the agent starts replying on its own.
## How do I know it worked?
**Both** conditions have to be true:
1. The agent shows as **Active**.
2. Under **Settings → Sessions** there's at least one **connected** session (not *Not connected*).
If either is missing, the agent won't reply.
## FAQ
No. It's **reactive**: it only replies to incoming messages, within the WhatsApp conversation window. To re-engage idle conversations, use [Automatic follow-up](/en/guias/seguimiento-automatico).
It's almost always the session. Check **Settings → Sessions**: if it says **Not connected**, the agent has no line to reply through.
**Development** is your draft: you test there without affecting anyone. When you click **Publish**, that content moves to **Production**, which is what your customers see.
The **prompt** is the "who you are and how you speak". The **policies** are the specific data and rules (deadlines, prices, conditions). Splitting them makes it far easier to update a rule without touching the agent's personality.
No. That feedback **creates a ticket** the Mindo team reviews. For an immediate change, edit the prompt or the policy yourself. See [Train your AI agent](/en/guias/entrenar-agente-ia).
Yes, with **Suggestion mode**: the agent drafts and you approve or reject. See [Suggestion mode](/en/guias/modo-sugerencia).
You can pause it per contact. See [Pause the agent permanently](/en/guias/pausar-agente-permanente).
Enable or disable your agent with a single switch.
Review every reply before the agent sends it.
Create your AI agent in the Mindo platform.
* **Tools**: actions the agent can run, like checking stock or an order status. Click **Add new tool** to browse the catalogue.
* **Guardrail**: enabled by default. It classifies greetings and goodbyes so the agent doesn't over-reply. Turn it off and the agent always answers.
* **Follow-ups**: automatic messages to re-engage conversations that went quiet.
* **AI Simulation**: create test customers with different personalities. You can **auto-generate them from the prompt** or **create them manually**.
* **Standard Tests**: load fixed questions with the answer you expect, click **Run All** and Mindo tells you which ones passed.
* **Reviews**: **Create Review** generates a review so your team can rate the agent's interactions.
### Quick Chat
The fastest test: you write as if you were a customer and see how it answers. Under each reply you get 👍 / 👎.
### Sessions and Audience
The **session** is the line the agent replies through. Click the **+**, tick the line and connect it. The **Audience** is optional and limits which contacts the agent responds to.
The **Full history** button in the header opens the complete log with every change together. It's the best place to understand why the agent started answering differently.
## Publish and turn it on
