What's in this article?

This article covers how to configure agents in AI Studio, with a focus on Conversational and Copilot agents. These two agent types are similar in all ways except for the deployment process.

After creating an agent, you will land on the Agent Configuration Wizard — a no-code flow designed to help you quickly get your agent configured, tested, and deployed to production. As you navigate through the configuration flow, saving the agent will make updates immediately available for testing in the Agent Testing panel. The one exception is when you add new Content sources. You will need to train your agent to see updates from new sources.

Agents also have versions. You will see Draft, Deployed, and Previously Deployed versions of the agent in the header. If you are on a Deployed version, clicking Edit as Draft will create a new Draft version of the agent, which can then be edited. You can only have one Draft at a time.

Details

The Details screen contains basic information about your agent: its name, description, and avatar. It also includes model selection, where you choose which LLM model you want the agent to use and which fallback model it should use if the primary model is unavailable. The model list is drawn from your Approved Models list in AI Studio > Models. Optionally, you can also add your organization name and brand voice to further customize the tone for this agent. Click Save and Next to continue.

Workflows

The Workflows screen has three main sections: Prompt, Rules, and Advanced Settings.

Primary Agent Prompt

This is the primary prompt your agent uses on every interaction. It tells the agent who it is and what its job is, and the agent uses it to accurately answer questions in the proper way. A default prompt is provided, but you are free to customize it to meet the specific needs of your agent.

Rules

Rules define specific guidelines that control how the AI agent behaves and responds. They help fine-tune the agent's behavior through reinforcement learning from human feedback (RLHF). Your agent will refer to these rules each time it generates a response. Keep them simple and focused so you do not introduce competing priorities for the agent to follow. Rules can set both what to do (for example, always talk like a pirate) and what not to do (for example, never use tables in responses). You can add rules and then activate or deactivate them without deleting them, which is useful for quick testing.

Advanced Settings

These are less frequently used settings that allow you to further refine your agent. For example, you can add restricted topics that the agent is not allowed to discuss, or adjust other model-level configurations.

Content

The Content Sources section is the primary data that your agent is trained on. You can add files, URLs, or text directly. All content is vectorized to allow the agent to quickly and easily find answers from your content sources when asked. This section should be used primarily for content that does not change often — such as policies, brand guidelines, and company information — since it will only be updated when you retrain and redeploy the agent. For content that changes frequently, such as a knowledge base, it is better to use a tool that searches that content each time it is queried so the agent is always retrieving the latest information.

Tools

Tools are the actions your agents can take. You have complete control over which tools each agent is allowed to call. Once you have approved a tool from AI Studio > Tools, it will appear in the Tools section of the agent configuration flow. You can then enable the tool for this agent and configure any relevant details on this screen.

This screen also displays your available Workflows. You can grant your agent access to those workflows here, and the workflow will trigger based on the trigger configured on the workflow itself. After enabling your tools and workflows, click Save and Next to test them in the Agent Testing panel. Note that the testing panel does not have the same context as a Copilot agent (for example, a case ID), so you will need to provide the data required to run the tool manually during testing.

Deployment

The Deployment screen is where you configure your agent for use by real users. It is good practice to always train and test your agent before deployment. Click Train and it will vectorize all your content sources. This process is quick for text and file sources, but if you have URL sources it can take longer — approximately 5 to 20 minutes — as it needs to scan and scrape all content from the provided URLs.

You can then run your experiments for this agent to ensure quality remains high after your latest updates. Once ready, you can choose where to deploy it. The deployment options differ between Copilot and Conversational agents, as described below.

Copilot

No additional action is needed for deployment. When deployed, your new Copilot Agent will be available for use on the Copilot tab on Cases. You can also optionally trigger this agent via API if you want to integrate it into your own application; the API endpoint is available on this screen.

Conversational

Conversational agents offer three deployment options: Voice Channel, Chat Channel, and Conversational AI Actions.

Testing Your Agent

You can test your agent draft directly in the configuration wizard without needing to deploy the agent. Just start a conversation on the Agent Testing panel on the right and you will see how it will respond. You agent will have all the latest configuration updates after each save and will have all your content sources after you train it.

You can also access the Trace Details page but clicking the icon in the bottom left of the footer of each message to see all the details of the agent thought process to get to its final response. You can also click on the thought process widget at the top of the response to see all the sources it found, the tools it used, and anything else it did to get to its final answer.