Skip to main content

Explore Conversation Design in the Agentic World

Learning Objectives

After completing this unit, you’ll be able to:

  • Explain the role of conversation design in creating effective conversational AI agents and experiences.
  • Apply the cooperative principle to improve interactions between humans and AI agents.
  • Identify and define the fundamental elements of conversation design.

What Is Conversation Design?

Conversation design (CxD) is the art and science of creating conversational experiences. Think of CxD as a blueprint for how a user interacts with a chatbot, a voice assistant, or any other AI agent. The goal is to make the interaction feel natural, useful, and successful. Good conversation design anticipates what a user says or asks, and it guides the AI to respond in a way that’s helpful and easy to understand. It makes sure that the conversation flows logically from one topic to the next, helping users accomplish their goals without confusion.

Are you wondering how designing a new AI agent, conversation design, and prompt design all fit together?

  • Conversation design creates the overall structure, user experience, and personality of an AI agent.
  • Prompt design provides the specific instructions the agent uses to fulfill its role within that structure.

This module focuses on conversation design, and touches a bit on prompt design.

What Makes a Good Conversation?

A good conversation is like a good partnership. Both participants work together to understand each other and achieve a shared goal.

Two people on a bench in a park having a conversation

[This image is AI-generated with Google Docs Gemini using this prompt: Two people having a conversation. They are sitting on a bench outside under a tree in a park. They are happy. Use the graphic style of Trailhead.]

In the field of conversation design, this partnership is guided by a concept called the cooperative principle. Developed by philosopher H.P. Grice, the cooperative principle suggests that, in any conversation, both participants implicitly agree to cooperate. To do this, they follow four main rules, or maxims.

  • Maxim of quality: Be truthful. Don’t say things you know are false or for which you lack evidence.
  • Maxim of quantity: Be informative. Provide just enough information, not too much and not too little.
  • Maxim of relation: Be relevant. Stay on topic and provide information that is pertinent to the conversation.
  • Maxim of manner: Be clear. Avoid ambiguity and be as brief and orderly as possible.

Apply the cooperative principle to everyday conversations that you have. For example, if your friend asks for directions, you don’t give them a history of the streets; you simply tell them the direct route. Those are the maxims of quantity and relation in action.

The same principle applies while designing conversations for an AI agent. You guide the agent to provide truthful, relevant, and clear information without overwhelming the user.

For example read these conversational snippets and learn the maxim that should be followed–but isn’t.

Dialog

Maxim Not Being Followed

Anna: "Did you eat the last brownie?"

Ben: "No." (Ben has chocolate on his face).

Quality

Chris: "What did you have for lunch?"

Diana: "I had a sandwich. The bread was whole wheat, toasted. It had turkey, which was sliced very thin, provolone cheese, a little bit of mayonnaise, and some lettuce that was getting a bit wilted."

Quantity

Eva: "It's getting late. We should probably head home."

Frank: "I wonder if dogs can see in color."

Relation

Greg: "Can you please pass the salt?"

Hanna: "It is within the realm of possibility that the object you desire could be transferred to your possession."

Manner

What Are the Fundamental Elements of Conversation Design?

To design a good agentic conversation, you need to understand its basic building blocks. Each piece of the conversation has a specific name and function.

  • Utterance: An utterance is what the user actually says or types. It's the literal text or spoken words. An utterance can be a question, like “What’s the weather today?”, a statement, like “I need to know the weather”, a command, like “Cancel,” or any other type of sentence, phrase, or word.
  • Response: The response is the AI agent’s reply to the user’s utterance. A good response is helpful, clear, and addresses the user’s intent. A good response is, “The weather today in Long Branch, Virginia is mostly sunny with a high of 75 degrees Fahrenheit.”
  • Intent: The intent is the goal or purpose behind a user utterance or AI agent response. It’s the why of the conversation. An intent can be expressed through many different utterances or responses. Examples include: “How's the weather?” or “Tell me if it’s going to rain.” These are different utterances, but they share the same intent: to get a weather forecast.
  • Turn: A turn is a single complete statement or action by either the user or the AI agent. The user’s utterance is one turn, and the agent’s response is the next.
  • Turn-taking: This describes the process of each participant taking their turn in a conversation. A smooth conversation has natural turn-taking. For example, a user asks a question (Turn 1), the agent responds (Turn 2), and the user asks a follow-up question (Turn 3).
  • Prompt: A prompt is the natural language instruction a designer gives to the AI agent to guide its responses. It’s how you tell the agent what information to include or the level of detail. For example, an instruction like, “When telling someone the current weather, include the city name, current weather conditions, and the temperature,” prompts the agent how to respond to weather inquiries.

Agentic Interactions

Conversational AI has evolved significantly. Understanding this evolution helps you design for different types of agents.

Deterministic Versus Generative Output

The core difference in how earlier AI chatbots responded and how newer AI agents respond lies in how an AI agent generates its response.

  • Deterministic output equals preprogrammed responses. Chatbots were only able to respond to user intents using a predetermined library of scripted responses. For each chatbot intent, there was one response or formula to use. No creativity or flexibility was available.
  • Generative output is created on the fly. An agent with generative capabilities uses a large language model (LLM) to create a new response each time. The agent doesn't just pull from a list of canned responses; it generates a new one based on the context and the utterance.

For example, consider the difference between a deterministic and a generative response for that question about the weather.

Deterministic Response

Generative Responses

The weather in [Location] is [Condition] and [CurrentTemp]

“Right now in Long Branch, it’s 75 and sunny”

“Currently it’s sunny and 75 degrees in Long Branch”

“It’s a beautiful day in Long Branch , 75 and sunny.”

Conversation Design for Traditional Chatbots

With a traditional chatbot, you wrote the exact responses for every possible user interaction. You anticipated all the different ways a user can ask a question, and then wrote a direct, predefined answer for each one. This was very structured and rule-based. It was like creating a series of if-then statements.

For example:

  • When a user said, “I need help with my password.”
  • The chatbot responded, “To reset your password, visit our help page.”

Conversation Design in an Agentic World

In an agentic world, your role shifts from writing the exact responses to guiding the agent on how to generate a response. Because of this, in the long run it’s easier to set up an agent than a bot. You provide the agent with a persona, context, and a set of prompts and instructions. You set the rules for what the agent must and mustn’t say. Instead of writing the specific response, you write the guidelines for the agent.

For example, you instruct your agent to:

  • Be a helpful customer service agent
  • Be polite and friendly
  • Not share personally identifiable information (PII)
  • Provide a brief, factual description about products

Your job is to define the context and guardrails for the AI, giving it the tools and rules to create its own responses that are still aligned with the brand and the user’s needs.

Next Steps

You’re well on your way to conquering the art of conversation design. In the next unit, you learn the framework and process to develop great conversations.

Resources

Share your Trailhead feedback over on Salesforce Help.

We'd love to hear about your experience with Trailhead - you can now access the new feedback form anytime from the Salesforce Help site.

Learn More Continue to Share Feedback