Agentic conversations in Virtual Agent
Summarize
Summary of Agentic conversations in Virtual Agent
Agentic conversations in Virtual Agent enable the virtual agent to understand complex user queries by reasoning, planning, and executing tasks across AI agents, skills, topics, knowledge base (KB) articles, and other supported resources. This capability allows Virtual Agent to orchestrate multiple agents and skills dynamically to fulfill user requests efficiently, improving response quality and task completion.
Show less
Enabling Agentic Conversations
To use agentic conversations, administrators must manually enable AI agents within Virtual Agent by:
- Creating and configuring multiple assistants with defined scopes.
- Assigning assistants to specific portals or channels.
- Enabling agentic support on the assistants.
- Mapping or publishing AI agents to assistants via AI Agent Studio.
Only configured agents within the current assistant scope are active during conversations, allowing the system to orchestrate multiple agents and skills to address user requests.
Examples of Agentic Behavior
- Multiple KB questions: Breaks down compound questions into individual queries, providing focused answers rather than a mixed response.
- Multiple skills with slot filling: Separates multi-part requests (e.g., ordering coffee and food) into distinct tasks and executes them sequentially.
- Complex multi-task requests: Plans and executes a series of dependent tasks across agents and skills, using outputs from prior actions to complete the entire request.
- Combination of QnA and agent actions: Recognizes multiple intents in one utterance and processes them in sequence, even if one intent involves knowledge retrieval and another involves an action.
Managing Conversations
Users can halt an ongoing agentic conversation mid-query by selecting an interrupt icon, which stops the conversation and allows restarting with a new query.
Error Handling and Fallbacks
The Virtual Agent provides messages and fallback options when encountering issues such as:
- Unrecognized user requests — options may include contacting a live agent, creating support tickets, or viewing help resources.
- Session inactivity — warnings prompt users to continue before timing out the session.
- Technical errors — messages inform users and offer alternative options to proceed.
- Failures during multi-step agentic tasks — the agent notifies users of the issue and suggests fallback options based on configuration.
Important: Always review AI-generated responses for accuracy before taking action to ensure reliable outcomes.
When you ask a question to the virtual agent, the agent understands the query. It can reason, plan, and execute across AI agents, virtual agent topics, conversational actions and subflows, catalogs, KB articles, custom skills, and any Now Assist in Virtual Agent supported skills to help you.
- If for the given assistant, specific agents are available to perform user tasks or sub tasks, they’re used.
- If a specific agent isn’t available for the task or sub task, the system automatically employs the Search Agent to discover answers or appropriate skills within the system (again based on the assistant scope).
- If skill execution is required, the system automatically executes the discovered skills.
- The system can plan and orchestrate execution among multiple agents, skills, and QnA to accomplish complex tasks.
Enable AI agents in Virtual Agent
Role required: admin or virtual_agent_admin
- Create and configure multiple assistants with specific scope and map the assistants to one or more portals.The configuration consists of the following:
- Creating an assistant in Virtual Agent or using the default. To create an assistant, see Create a chat assistant
- Assigning specific assistants to a specific portal or portals. For more information, see Display your chat assistant on a portal, channel, or mobile app.
- Ensure that Agentic support is enabled. For more information about using agentic support, see Use agentic support for a chat assistant.
- Map or publish an agent to one or more assistants on AI Agent Studio to make the agent available within a specific assistant. For more information, see Create an AI agent.
During execution, only the configured AI agents are considered for the current assistant and dynamically makes them available to the Orchestrator for planning.
Examples of AI agent behavior for user utterances
| Agents | Skills/Topics | KB Articles |
|---|---|---|
|
Check IT Ticket status agent Email Agent Meeting scheduling Agent |
Order coffee Order food Order laptop Order accessories |
Spam ESPP policy |
Scenario 1: Multiple questions from KB articles
Utterance: How do I avoid spam? How do I detect it?
- Non-agentic response: Produces a single mixed answer.
- Agentic response: Breaks it into two questions and provides a better answer for each one of them.
Scenario 2: Multiple skills with slot filling
Utterance: Hey, order a coffee for me, preferably dark roast and something to eat, maybe a pizza?
- Non-agentic response: Produces a single answer. Mostly listing all matching available skills. No auto-execution since it matched multiple skills.
- Agentic response: Breaks it into two distinct tasks, order coffee and order food/pizza. Executes one after another, completing the entire user request.
Scenario 3: Complex utterance with a combination of skills, agents, and QnA
Utterance: I am going on PTO tomorrow. Get my expense report and my IT ticket status. Send a summary of the expense report to John Jacob and the details on ticket status to Robert Williams, informing them of my PTO and requesting them to work on them.
- Non-agentic response: Produces a single answer. It lists all matching available skills. No auto-execution will take place since it matched multiple skills.
- Agentic response: Breaks it into multiple distinct tasks, reasons and plans, understands the dependencies, and executes one after another, completing the entire user request using output from prior actions as context as needed.
Scenario 4: Complex utterance with a combination of QnA (KB) and agent
Utterance: What is the maximum contribution amount for espp? Send an email to Robert with the details.
- Non-agentic response: Produces a single answer and doesn’t complete or even suggest the second action since there’s no corresponding skill.
- Agentic response: Understands the two separate intents and executes them in sequence while using the output from the first intent to fulfill the second intent.
Halting and restarting agentic conversations
If you want to stop an agentic conversation mid-query, hover over the send icon 
while the agent is researching a query. The icon becomes an interrupt flow icon 
. Select the icon and the conversation stops. A message appears: The current conversation has been stopped, but you can begin again. Enter a new query to restart the agentic conversation.
Error handling and fallback options
When the AI agent cannot understand a request, encounters a system error, or a session becomes inactive, it displays a message and presents options to help you continue.
- When Now Assist cannot find an answer:Depending on how your administrator has configured the assistant, these options may include:
- Requesting a live agent
- Creating a support request
- Viewing office hours
For example, if you ask about a topic the AI agent cannot find information on, you might see a message such as: “I can’t find anything on that directly but I can help you explore other options.” Select one of the displayed options to continue.
- Session timeout due to inactivity:
If you stop responding during a conversation, the AI agent waits for a set period before ending the session. When a session is about to time out, you will receive a warning message. For example: “Looks like you’ve left this conversation. Would you like to keep going?”. If you want to continue, respond to resume the conversation.
If there is no response to the warning, the session ends with a session end message. For example, “Our chat has timed out due to inactivity. Just start a new conversation whenever you’re ready. I’ll be here to help!”. You can start a new chat to continue getting help.
- Technical errors and system timeout:
If the AI agent encounters a backend error or cannot complete a request due to a technical issue, it displays a message and presents fallback options. For example: “It looks like I’m having trouble completing your request right now because of a technical failure. But I can help you explore other options.” You can select one of the available fallback options to continue.
- Errors during agentic tasks:
When you ask the AI agent to complete a multi-step task on your behalf, it plans and carries out the steps automatically. If a step in the plan cannot be completed, it lets you know and offers fallback options so you can decide how to proceed.
For example, if the AI agent is creating an expense report and encounters a problem, it displays a message such as: “Unfortunately that didn’t work. But I can help you explore other options.” The fallback options available depend on your configuration.