# Getting started with agents in datasets

> Learn how to build multi-tool agents with structured outputs that iterate through your data to research, analyze, and report.

Source: https://cotera.co/docs/guides/datasets/agents-in-datasets

---

LLM agents in Cotera become significantly more powerful when you chain multiple tools together. While a single tool can fetch data, a multi-tool agent can act as a complete research loop: searching for information, visiting websites to read content, analyzing the findings, and returning data in a specific, machine-readable format.

***

[Video: Getting started with agents in datasets tutorial](https://youtu.be/xzDjSTiZZjM)

***

## What You'll Build

In this guide, you will learn the architecture of a "Research, Analyze, and Report" agent. You will build an agent that:

* Iterates through a dataset (like a list of cities)
* Uses search and scraping tools to find specific answers
* Takes action (like sending an email)
* Saves the result in a strict, structured format

***

## Prerequisites

Before starting, make sure you have:

* A Cotera account with access to agents and tools
* A dataset to iterate over (e.g., a CSV list of cities, companies, or topics)

***

## Step 1: Initialize the Agent

Navigate to your dataset and prepare the column that will house your agent.

1. Click the **+** button on your table to add a new column
2. Select **LLM Agent** from the menu
3. **Name Your Agent**: Give it a descriptive name (e.g., "Market Researcher")
4. **Select a Model**: Choose a capable model (e.g., GPT-4o or GPT-5) that can handle complex reasoning and multiple tool steps

***

## Step 2: Finding and Installing Tools

Before writing your prompt, you need to ensure your agent has the right equipment. Tools come in two categories: those already in your workspace and those you need to install.

To add a new tool (like a web scraper):

1. Click the **Install Tools** button (or the tools menu icon) in the agent interface
2. Search for the functionality you need (e.g., "Scrape")
3. Select the tool, such as **Cotera Built-in Enrichments / Scrape Website**
4. Click to add it—this often provides a specific "handle" or code snippet to use in your prompt

***

## Step 3: Crafting the Prompt

The System Message is the brain of your agent. You will build this prompt layer by layer.

### Define the Persona and Variables

Start by telling the agent what it is. To link the agent to your data, type **** to see your list of columns:

```
You are a research agent. I will give you a {City} and {State}.
```

> The variable names will highlight in yellow when linked correctly to your dataset columns.

### Reference Your Tools

To tell the agent to use a tool, type **@**. This opens the tool dropdown:

```
First, use the @Google Search tool to find the "best ice cream" in this location.
```

### Chain the Logic

Explain how the output of one tool should become the input of the next:

```
Once you have the search results, use the @Scrape Website tool to visit the top URLs and read the reviews.
```

### Add an Action (Optional)

You can instruct the agent to communicate outside of Cotera:

```
Finally, use the @Email tool to send a summary to my.email@example.com.
```

***

## Step 4: Structuring the Output

One of Cotera's most powerful features is forcing the LLM to return data in a specific computer-readable format, rather than just chatty text. This makes your data sortable and filterable.

To define your output, click the **** **** icon to open the output structure menu.

You can select from several data structures:

| Type                 | Best For                                                                                                              |
| -------------------- | --------------------------------------------------------------------------------------------------------------------- |
| **String**           | Summaries, emails, or free-form text answers                                                                          |
| **Boolean**          | Binary decisions (e.g., "Is this company hiring? True/False")                                                         |
| **Enum**             | Forcing the agent to choose from a pre-set list of options (e.g., "High Priority," "Medium Priority," "Low Priority") |
| **Object**           | Structured JSON objects with defined keys (e.g., shop\_name, rating)                                                  |
| **Array of Objects** | A list of structured items (e.g., a list of 5 different products found on a single page)                              |

**Example**: If you want the agent to return the name of the best business found, select **String**. If you want it to return the business name and its address, select **Object** and define those fields.

***

## Step 5: Save and Run

1. **Save**: Click the **Save** button in the bottom right
2. **Run**: Click **Run Now** on the column

### Monitoring the Agent

As the agent processes your dataset, you can watch it think in real-time. You will see status updates in the cell indicating which tool is currently active:

* Thinking...
* Searching Google...
* Scraping website...
* Sending email...

***

## What You've Built

You have moved beyond simple text generation. By combining tools (external access) with structured outputs, you have built an autonomous research agent that:

* Reads your database
* Navigates the live internet
* Performs actions
* Returns clean data ready for analysis