Design intelligent conversations in Copilot Studio

In the world of conversational automation, Microsoft Copilot Studio offers a platform for building intelligent agents that interact with users naturally and efficiently. In this article, we explore the concept of generative orchestration where topics with a ‘By agent’ trigger come be used, view their role in the conversational architecture, and how developers can leverage them to create business solutions.

What is generative orchestration in Copilot Studio?

Generative orchestration allows an agent not only to answer questions and follow a defined flow (classic orchestration) but also to perform complex actions, enabling the agent to dynamically decide what action to take based on the user’s context, intentions, and the available data.

What are topics with an ‘Agent’ type of trigger?

Topics in Copilot Studio define how an agent responds to a user’s intent. Each topic is triggered by trigger phrases and contains a series of steps that may include messages, conditions, actions, and branches.

Essentially, a topic is like a function in programming: it encapsulates specific logic that can be invoked when a particular intent is detected.

Agent-triggered topics are an evolution of the traditional keyword-based model. In these, the agent acts as an intelligent agent that detects the user’s intent and automatically activates the most relevant topic.

How do you set it up in Copilot Studio?

First, access the agent we are designing in Copilot Studio, then in Overview tab, we select the AI model that our agent will use.

In Settings / Generative AI, enable the Orchestration option. This feature allows the use of generative AI within topics, enabling the agent to make intelligent decisions based on the conversational context.

Once activated, we will see that we can use topics whose trigger is ‘By agent,’ which indicates that they will be automatically activated according to the user’s intent.

To define the agent’s behavior, it is essential to properly define the agent’s context. To do this, in the Instructions section of the General Information tab, we must write in detail how the agent should behave, what type of commands it should follow, and describe to it some example cases. This configuration directly affects the quality of the responses and the accuracy of topic activation.

Finally, it is necessary to clearly describe in the topic’s details the conditions under which it should be used, for the agent to know when to execute each topic. The more information and examples we provide, better results will be in intent detection and in the automatic activation of the appropriate topics.