# Building AI Agents powered by MadKudu

<figure><img src="https://1835787252-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FV8CtRF59w8oxkhPMGwwB%2Fuploads%2F5TMk9zyheYVk8PeNhbdW%2Fimage.png?alt=media&#x26;token=4a2eada5-02b3-43f8-a926-8f2eea0bd109" alt=""><figcaption></figcaption></figure>

<figure><img src="https://1835787252-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FV8CtRF59w8oxkhPMGwwB%2Fuploads%2FU8nIX19N9yH9YVAfHsB1%2Fimage.png?alt=media&#x26;token=b08a1138-5df7-4662-9279-92b3efbf6cdf" alt=""><figcaption></figcaption></figure>

## 🤖 What is an AI Agent?

An **AI agent** is an automation with AI in it.&#x20;

In practice, this means there is one or more parts of it where an LLM **makes a decision** on its own instead of having only determinist "IF this THEN that" as rules. &#x20;

AI agents introduce a level of **autonomy and reasoning.** They can act on ambiguous instructions (e.g. “find the best contacts to reach out to”) and pulling in data from various systems (like MadKudu) to make decisions.

> Example: an AI agent drafting emails decide on its own what to write in this email. While a non-AI workflow would use a template text with placeholders and the email would look similar for all recipients.

AI agents are not products—they are **custom-built automations**. And like any automation, when something goes wrong, it’s important to trace which part of the system is responsible

## 🧱 The 4 Components of Any AI Agent

To build a working AI agent using MadKudu data, you need **three components**:

### **1. The Orchestration Platform**

This is **where the agent logic lives and gets executed.** It’s the engine that chains steps together.

Example: n8n, Make, Zapier, Dust, custom code running in Cursor ... \
These platforms execute the flow: pulling data, making decisions, sending messages, etc.&#x20;

### **2. The LLM**&#x20;

This is the **“thinking” layer** of your agent.

You can use: OpenAI, Claude, Mistral models, any model you want&#x20;

### **3. The Prompt**

The prompt contains the instructions in plain english that will be interpreted by the LLM&#x20;

Example:&#x20;

* "Find contacts at {company}"
* "Draft an email for {email} using all the information you know from MadKudu"&#x20;

{% hint style="warning" %}
LLMs are powerful but inherently unpredictable. They can hallucinate, misinterpret instructions, or fail entirely when given poorly structured inputs. When building agents, much of the effort goes into **refining prompts** and **selecting the right model** to minimize the risk of misinterpretation and ensure more consistent and reliable outputs.
{% endhint %}

### **4. The Data layer: MadKudu MCP**

MadKudu provides the **data layer**—the context that inform the agent’s decisions.

Through the MadKudu Public API and MCP integration, you can&#x20;

* get information about your accounts and contacts: firmographics, demographics, scores, engagement activities with your website, product, marketing, job history...&#x20;
* search for accounts and contacts&#x20;
* source and enrich net new contacts&#x20;

## 🧪 Example Use Case: Meeting Prep Agent

Here’s how a MadKudu-powered AI agent for meeting prep might be built:

* **Orchestration**: n8n monitors your calendar, detects meetings with prospects.
* **LLM**: GPT-4 summarizes MadKudu data and past emails into a briefing doc.
* **MadKudu**: The MCP provides the account brief, top users, recent activity, and firmographics.

You decide when the agent runs, what it generates, and where it shows up (Notion, Slack, Google Docs…).

{% hint style="success" %}
MadKudu helps power your AI agents by exposing real-time GTM data through secure APIs and the Model Context Protocol (MCP). But we don’t control:

* Which tool you use to orchestrate
* How you write your prompts
* How your model interprets the data

Treat MadKudu as your **intelligent data layer**—not the AI agent itself.
{% endhint %}