Before you start

Prepare at least one Supsearch search engine with trained content. Agentic nodes use Supsearch as their knowledge layer.

Recommended first build scope

For a first rollout, build a single-use-case agent that answers FAQ-style questions from one knowledge source and hands over edge cases to a human. The video walkthrough follows this pattern using a single Agentic node connected below a Greet node and a single knowledge source.

Step-by-step: create your first agentic virtual agent

1. Create a new virtual agent
 

1. From Manage Virtual Agents, click Create New Agent.
2. Enter a clear virtual agent name.
3. Select the language.
4. Choose the channel. In the walkthrough, the agent is created for Chat.
5. Click Create to open the BotStudio graph with a default Greet node.

2. Open Conversation Flow and add a child node
 

6. In the graph, start from the Greet node.
7. Create a new child node beneath Greet.
8. Name it clearly, such as “Agentic”, “Broadband assistant”, or the use case name.
9. Select Agentic Node as the node type.
10. Open the new node to start the guided configuration on the right-hand panel.

3. Configure knowledge
 

11. Add a search engine from the dropdown.
12. Write a description that tells the model when to use this knowledge source. In the video example, the description is effectively “use this search engine when answering questions about Puzzel Broadband”.
13. Optionally add tags if you want to restrict which articles can be used.
14. Set a search confidence threshold only if you want to filter out weaker search results. Articles below the threshold are ignored.
15. Leave “Show links used for answer generation in agentic responses” enabled if you want source links returned to the user.

4. Configure exceptions
 

16. Review the default handover exception.
17. Add exceptions for cases where the agent should stop being autonomous and route to a deterministic flow or human queue.
18. Good starter exceptions are: speak to a human, order issue, complaint, identity verification, or any workflow that needs backend integration.
19. Each exception needs a name, a prompt describing when it should trigger, and a child node to route to.

5. Add guidance
 

20. Use guidance to define the agent’s role, tone, output style, and hard rules.
21. Decide whether to inherit global instructions from Configuration > Generative AI. This is enabled by default.
22. In the walkthrough, the guidance follows a grounded-answer pattern: always search for articles first, answer only from the retrieved content, politely decline when the answer is not found, and respond politely to pleasantries or clarifying questions.
23. Keep first-version guidance short and explicit. Long prompts are harder to tune.

6. Set temperature

24. Use a lower temperature for factual support scenarios.
25. The recorded example uses a low setting around 0.3, which is a good starting point for customer support use cases.
26. Increase temperature only if you need more expressive phrasing and variation.

7. Configure guards and fallback
 

27. Choose whether to enable Answer Relevance, Groundedness, and Word Similarity.
28. Keep guard settings conservative at the start. Overly strict settings can prevent the node from answering at all.
29. Define a custom fallback node if you want behavior other than the Global Fallback Node, such as asking the user to rephrase or routing to human handover.

8. Finish configuration

30. Click Finish to save the Agentic node.
31. Confirm the node is connected correctly in the flow and appears under the Greet node or whichever parent node you selected.

9. Test the experience

32. Use the demo panel on the right side of BotStudio to ask representative user questions.
33. Enable status messages in Configuration > General to see what the node is doing during a conversation.
34. Use the footsteps icon to inspect logs for exception classification, knowledge search, guard evaluation, and final output generation.

10. Tune after testing

35. If answers are weak, improve the knowledge source before rewriting prompts.
36. If the agent answers when it should route, tighten exceptions.
37. If the agent is too rigid, soften the guidance or reduce guard strictness.
38. If hallucinations appear, lower temperature and strengthen grounding instructions.

Configuration cheat sheet

What to expect in your first release

• Expect the quality of answers to depend heavily on knowledge quality. A well-structured search engine matters more than clever prompting.

• Expect to iterate on exceptions. Early versions usually either over-route or under-route.

• Expect some tuning of guidance, temperature, and fallback behavior before the experience feels production-ready.

• Do not expect the Agentic node to replace every deterministic flow. Use it where flexible understanding helps, and keep fixed flows for transactions, compliance-heavy journeys, and sensitive workflows.

Common launch mistakes to avoid

• Starting with too many knowledge sources and no clear descriptions.

• Writing very long guidance instead of a few high-priority rules.

• Skipping exception design and relying on the model for everything.

• Turning guards up too high before understanding baseline answer behavior.

• Testing only happy-path questions instead of real customer phrasing.