Virtual Agent with agentic reasoning in the October release, and what it means for you

[Victor Chen]

The October 16th release for Now Assist in Virtual Agent brings changes that will affect customers who upgrade to Yokohama Patch 8 or Zurich Patch 2.

• If customers do not upgrade to the latest patches, there are no changes. 

• If a customer upgrades to YP8/ZP2 and the latest corresponding plugins for Now Assist in Virtual Agent (v12.x) and GAIC (v11.2), and if the customer has the “AI Agents” skill checked for their assistant setup, you will expect the following changes as part of “Virtual Agent with agentic reasoning":
• The default LLM model is switching to GPT-4.1. Customers who use NowLLM will switch to using GPT-OSS. 
• Multi-intent is supported and will appear in responses
• AI Agents are prioritized in Virtual Agent responses by default, but this is configurable via a system property - sn_aia.use_agents_in_planner
• Improved performance and latency - see the stats and video below!
  • ~20-41% faster synthesized response speed with 3rd party models
  • ~40-52% faster synthesized response speed with NowLLM
  • ~27% reduction in E2E run time for OOB agentic workflows and AI Agents
  • More improvements expected in Q4 release and beyond [Safe Harbor]
• UI updates to in-chat processing messages

• Customers who enable Virtual Agent with agentic reasoning will not be able to turn it back off. If you do not wish to automatically uptake these changes when upgrading, please uncheck the “AI agents” skill in assistant setup before upgrading (via Assistant > Now Assist skills). 

The full KB for these changes can be found here (for customers): KB2548818.

Now Assist in Virtual Agent - speed comparison video

Screenshot of multi-intent in action

[qprinsloo]

@Victor Chen , the KB article you've linked (KB2548818) is restricted and cannot be viewed even after logging in to support.

[Victor Chen]

@qprinsloo - try again? (You must be a customer in Now Support to view this KB.)

[David Schafer]

I'm logged in as a customer...

[anand9]

I logged in, but KB is still not visible. showing error "Your role does not grant you access to this article. Please use Search to find related content."

[Victor Chen]

We'll look into this. In the meantime, the article has been updated with specific performance improvement stats!

[LohitYogi]

@anand9 can you please try again? 

[anand9]

@LohitYogi - same error:- 

[anand9]

What are the other changes in Zurich patch 2. My custom Agentic AI had a accuracy of 80% on patch 1, which is significantly dropped to <50% immediately after moving to patch 2, and I am kinda frustrated having trust issues (instead of having better performance and accuracy, I am getting issues and have no idea where to look) - I am not using VA yet. Now sure what was the changes in patch 2, but How do I revert to what was there in patch 1? 

[himmat10005]

I am also not having access to this KB Article although I am logged-in with my customer account.

[Eric Merkle]

@anand9 can you share more about the details of what is "My custom Agentic AI" and if you submitted a case on the issue?  would love to hear more on the challenges and how we can improve .   

[anand9]

@Eric Merkle - I am developing a custom Agentic AI, where good amount of the AI instructions are being picked from article (as it is not a good idea to keep a lengthy instruction in workflow, and also it gives a customer chance to modify the instruction later if required).

The Agentic AI is related to CSDM, where I am defining some definitions in article. For example, in Article, I have mentioned Application service, service instance and deployed instance are the same thing.

Zurich Patch 1 - when user wants anything to do with service instance, or deployed instance, Agentic AI would pick the right AI Agent and execute the workflow (most of time).

Zurich Patch 2 - Immediately after upgrading my instance to patch 2 - same question from user on service instance, or deployed instance, is triggering irrelevant AI Agent.

I have been following "AI Agents Prompting Guide" uploaded by Victor few weeks back - and I am sure there's nothing wrong with instruction, and yet, I am helpless why workflow is picking wrong AI Agent. Nothing valuable I see in Decision log as well. It was almost working before upgrade, now I am afraid to do any changes in instruction, as it would break something else.

Did I raise ticket - NO - because I have raised similar ticket (CS8200183) in past, and it was straight rejected due to the fact that it was a "custom" AI Agent, even though the case agent agreed that the provided instructions are good, but system behavior is not. So, I am sure they will reject it again. 

There are other non-consistent issues (occurred before patch upgrade as well) in how AI Agent works, but for now I am trying to fix the issue occurred after the upgrade. 

[Eric Merkle]

so the discovery of the AI agent with the same intent now seems to be a issue from ZP1 to ZP2 ? I'll share with the team 

[sriii1]

@anand9 Could you pls create a case, it would be hard for me to understand what your end to end scenario is and what is not working upgrading to ZP2. 

We can get more context through case, like you have upgraded to ZP2 and Oct version of store apps as well. Are you testing through VA, NAP or playground in Agentic Studio etc. How you are developing agents that picks up instructions from KB, how are you making those available to Agentic platform. Why can't you put instructions in worlflow/agent itself, anyone should be able to change those by going into workflow/agents. 

[anand9]

@Eric Merkle , @sriii1 - created CS8662226 with required details.

[Eric Merkle]

thanks the team will look into it 

[MikeJ1]

The KB (KB2548818) still isn't available will this be resolved or is there a different KB link

[ayman_h]

Hi,

The latest update definitely brings in faster performance but we noticed the conversational experience where there follow-up questions if VA is unable to understand the intent of the query is now gone.

An example prompt: I want to install software

Before this release: There would be follow-up questions to verify which software the user would want to install 

After this release: We noticed the bot would try to answer the question straight-away

Is there a way to decrease the 'temperature' to make it ask follow-up questions to understand the intent before responding? 

[Mitch Fleischm1]

Hi @anand9, @MikeJ1, @qprinsloo, @himmat10005, and everyone else who might be reading this!

I think the knowledge article @Victor Chen linked to was hard-coded to a specific version (sys_id) of the KB, and it's probably been revised a few times since - so the previous version is no-longer available. Here's a link to the latest version of the knowledge article using the number: KB2548818. At least it worked when I tried it! (3:39pm Eastern time on 17 Nov, 2025). For reference, I am a customer with access to ServiceNow Support.

Hope this helps!

Mitch

[Victor Chen]

Thanks @Mitch Fleischm1 - I've likewise updated the URL in the article.