# Register an AI Agent

## 1. Create Agent

* To register an AI Agent into the TrustLogix system, you can navigate to the **AI Agents** screen using the top left menu. 
* Here, you can click **Create Agent**, and provide name, description, and status.

{% hint style="info" %}
Only AI Agents set to the **"ACTIVE"** status will be available in the policy engine
{% endhint %}

## 2. Configure Agent

* Once created, you **MUST** set up the agent configuration. Navigate to the created AI Agent, then to the **Configuration** tab. 
* Here, you have the option to configure the **Entitlement Source** and **OAuth**.

{% hint style="warning" %}
Currently, only other registered data accounts are supported as entitlement sources. Direct entitlement sync from configurable sources is a planned feature.
{% endhint %}

* OAuth configuration corresponds to the **AI Agent** identity. 
* We currently recommend using an existing registered application, as this identity will be used for agent activity monitoring in other registered systems in TrustLogix as well as **Just-In-Time** policies. 

{% hint style="warning" %}
Ensure the provided scope follows **Client-Credential/Self-Issued** scope formats according to your Identity Provider (i.e. Azure uses 'api://{client-id}/.default', Okta uses 'default', etc)
{% endhint %}

{% hint style="info" %}
If connecting to **Databricks**, select "**OAuth User to Machine Per User**" for Auth type. 
{% endhint %}

* After completing the form and clicking **Save**, an optional **API-Key** is generated, which can be used for additional security if securing a **custom** AI Agent by including it in the request headers of the MCP Client. 

## 3. Connecting Agent

* To communicate with AI Agents, TrustAI utilizes the **MCP** protocol to offer real-time policy decisions with respect to user identity, agent identity, and a host of conditions.
* See [mcp-server](https://docs.trustlogix.io/trust-ai-private-preview/mcp-server "mention") for more details. 

{% hint style="warning" %}
A common concern with connecting via MCP is the non-determinism in whether a policy is evaluated, as the AI Agent must call the tool. \
\
We recommend distributing responsibilities across your agents and orchestrating according to responses from the TrustAI MCP Server.\
\
If managing deployment yourself (i.e. LangSmith), LLM guardrails must be implemented by developers. 
{% endhint %}