AI Agents
CAI, Agents are the core building blocks of an intelligent conversational experience. Each agent is a finely-tuned entity designed to excel in a specific domain, equipped with unique knowledge bases, predefined Action capabilities, tailored communication style, and security protocols.
These specialized agents are not generic conversational interfaces, but precision instruments crafted to solve complex problems within their designated expertise—whether it's technical support, sales consultation, customer service, or strategic analysis. By combining deep domain knowledge with advanced reasoning capabilities, specialized agents can break down intricate challenges, generate context-aware solutions, and interact with users in a manner that mimics expert human professionals.
New Agent
Before creating your agent, you must have already added at least one skill (Action or Knowledge) to your Project. To do so, go to the respective sections and create the skills. Only then you’ll be able to create an Agent. Once you have added at least one, it’s time to build your team of agents.
Being goal-oriented, AI Agents possess a reasoning capability that allows it to execute actions based on information provided by users. The following fields shape how your agent perceives itself and its responsibilities, influencing every interaction and decision it makes.
Creating a new agent involves these three steps:
General Info
Role
A well-defined Role creates clarity about what users can expect from the agent and establishes the framework for all its capabilities, whether it's serving as a customer support specialist, a technical troubleshooter, or a creative assistant. Let’s walk through the creation of the agent Voyage Advisor.
Persona
You can either use the Project default persona or choose a specific one for each agent. You can use different ones for each of the agents depending of the use case.
Goal
The Goal articulates the primary objective your agent strives to achieve through its interactions. This strategic directive guides the agent's decision-making process, helping it prioritize actions and generate responses that consistently work toward fulfilling its purpose.
A clearly defined Goal helps your Supervisor agent to make intelligent choices when faced with ambiguity, ensuring that every conversation advances toward the right Agent.
Your Goal should be specific and detailed enough to clearly define the agent's expertise and capabilities. Vague or overly generic Goals (like "help users" or "provide support") make it difficult for the Supervisor to route requests correctly.
Your Goal should include:
Specific tasks and responsibilities the agent can handle
Key capabilities and services it provides
Domain expertise or specialized knowledge areas
Clear boundaries of what the agent does and doesn't do
Technical Support agent goal example:
Identify the user’s device and issue, perform a structured connectivity diagnostic, apply appropriate solutions, and confirm resolution or escalate the case according to operational rules.
Voyage Advisor agent goal example:
This agent suggests destinations to the user and creates itineraries based on the user's preferences, handles booking inquiries, provides travel recommendations for accommodations and activities, and assists with travel planning logistics.
Notice how it's specific about what the agent does (helps with technical issue, suggests destinations, creates itineraries, etc.) rather than being generic. This clarity helps the Supervisor understand exactly when to delegate travel-related requests to this agent.
The more specific and comprehensive the goal, the better the Supervisor can route appropriate requests.
Instructions
While the role defines what the agent is, and the goal defines what it should achieve, the instructions field defines how the agent should operate, behave, and execute its tasks. It provides step-by-step operational guidance and any specific rules that the Agent must follow during execution.
These instructions help ensure:
Consistency in the Agent’s responses
Compliance with process requirements
Proper sequencing of actions when interacting with the Supervisor or other Agents
✅ What to include
When writing Agent instructions, focus on clear, actionable steps and specific constraints relevant to the task. Here’s what you can include:
Step-by-step execution
Number the steps to make them easy to follow. Example: 1. Verify if fetch_invoice action has been executed at least once. 2. If not, execute fetch_invoice before issue_invoice. 3. Confirm with the user before proceeding to the next step.
Output formatting rules
Set character limits for each message.
Define the structure of the reply (e.g., summary + next step, list, table).
Specify if messages should be sent one at a time or combined all at once.
Protocols and compliance rules
Any mandatory procedures before moving forward.
Data collection requirements (e.g., "Always ask for the user’s email before sending a confirmation.")
Closing behavior
Define how the Agent should finalize the conversation for this use case. Example: "After completing the task, ask: 'Can I help you with anything else?' and, if not, prompt for feedback."
Fixed messages: You can set up standard and/or mandatory communications.
Enclose any fixed messages in quotation marks "like this" to ensure they are always delivered exactly as written. Example: If the user says they no longer need help, respond with: "Please rate this service at the end of our interaction."
When starting say: "Hello! I'm going to help you with
[process]. First, I need some information..." Upon completion, say: “Process completed! Is there anything else I can help you with?” For transitions: “I will now[next_action]. Please wait...”
❌ What NOT to include
Persona or tone of voice: These are already defined in the Agent’s persona or system-level persona.
General task description: This belongs in the goal field.
Ambiguous or contradictory rules: Avoid conflicting steps or unclear conditions.
Unnecessary context: Keep it relevant to execution, not high-level strategy.
Prompt example for a Technical Support agent:
You are a Technical Support Agent. Your goal is to diagnose and resolve device connectivity issues following the workflow below. Do NOT deviate from these instructions, and do NOT add extra explanations or commentary. Always output in HTML format.
Workflow steps (execute in this exact order):
Collect device type, operating system, issue description, and when the problem started.
Perform basic connectivity diagnostic, waiting for the user to complete each test before proceeding.
Apply the appropriate solution based on diagnostic results.
Verify issue resolution with user confirmation.
Document solution or escalate if unresolved.
Response formatting:
Output must be in HTML; do NOT use Markdown.
Do NOT include numbers or bullet points in messages.
Bold user actions using tags.
Limit each message to 200 characters.
Send one instruction per message; do not combine steps.
After each step, always ask: "Did this step work?" before proceeding.
Mandatory requirements:
Confirm device type before suggesting any solution.
MUST complete connectivity test before suggesting network solutions.
CANNOT recommend device reset until data backup is confirmed.
REQUIRED: Verify current settings before applying configuration changes.
Conditional responses:
IF basic solutions fail → Escalate to advanced diagnostics.
IF multiple devices are affected → Check for network-wide issues.
IF the problem persists after 3 solutions → Escalate to a human technician.
Fixed closing message:
When the user no longer needs assistance, output exactly: "Please rate my assistance at the end of this chat."
Important notes:
NEVER improvise steps or change their order.
NEVER output additional explanations outside the HTML-formatted instructions.
Preferences
Inherit Supervisor’s Preferences
You can choose to have Agents inherit the Supervisor’s preferences such as Guardrails and Constraints. Or, you can disable this option and override them with agent-specific tailored prompts to specialized functions.
Skills
Add Skills (resources such as Actions and Knowledge) that are already available in your project.
Agent cell
Once you click Save, the Agent cell, exclusive to Agentic and Combined models (not available in Intent-based/NLU models), will appear in the workspace.
The Agent cell is the main executor responsible of generating responses based on its goal, instructions, persona, and guardrails, after receiving a user input. If the case of the agent is not able to effectively respond to a user query, the conversation automatically transfers to a supervisor for further assistance.
For specific knowledge and tasks, the Agent cell can trigger skills like Knowledge sources or execute actions linked to it. When an action is associated with this Agent, a new branch appears in the workspace specifically for configuring that action's parameters.
The Agent cell supports any subsequent cell types except those incompatible with the agentic model (intents, entities, and end cells).
Last updated
Was this helpful?