Genesys Cloud - Main

Christoph,

 We are facing a similar issue, we had been conditionally rolled back to a previous opus on our org due to unexpected issues when Genesys updated the haiku model they were using to support freeform slots. I anticipate the v2 agentic stuff is in a similar boat. We've had to kind of scramble to re-test, but it appears it's less something Genesys has done and more with v3 being retired. 

Claude 3 Haiku - Amazon Bedrock

Best of luck with the testing! Hopefully you can employ an automated testing tool to decrease the lift.

Hi Jason,

thanks for your reply. An automated testing tool is on my Christmas wish list now. Maybe Genesys will provide something like that in the future. The integrated Testing tool from parloa is impressive.

Best regards
Christoph

Christoph,

 We are facing a similar issue, we had been conditionally rolled back to a previous opus on our org due to unexpected issues when Genesys updated the haiku model they were using to support freeform slots. I anticipate the v2 agentic stuff is in a similar boat. We've had to kind of scramble to re-test, but it appears it's less something Genesys has done and more with v3 being retired. 

Claude 3 Haiku - Amazon Bedrock

Best of luck with the testing! Hopefully you can employ an automated testing tool to decrease the lift.

We had an earlier preview of the APT-2 model and from my experience we're having to do a fair bit of rework to make it behave as it did with APT-1. in terms of testing, Genesys released Spec Driven Development for AVA which may be worth a look at. 

https://github.com/purecloudlabs/genesys-ava-skills

Hi Jason,

thanks for your reply. An automated testing tool is on my Christmas wish list now. Maybe Genesys will provide something like that in the future. The integrated Testing tool from parloa is impressive.

Best regards
Christoph

Christoph,

 We are facing a similar issue, we had been conditionally rolled back to a previous opus on our org due to unexpected issues when Genesys updated the haiku model they were using to support freeform slots. I anticipate the v2 agentic stuff is in a similar boat. We've had to kind of scramble to re-test, but it appears it's less something Genesys has done and more with v3 being retired. 

Claude 3 Haiku - Amazon Bedrock

Best of luck with the testing! Hopefully you can employ an automated testing tool to decrease the lift.

Hi Savino,

thanks for your insights. Good look with the change.

Thanks for the link, too. I will take a deeper look.

We had an earlier preview of the APT-2 model and from my experience we're having to do a fair bit of rework to make it behave as it did with APT-1. in terms of testing, Genesys released Spec Driven Development for AVA which may be worth a look at. 

https://github.com/purecloudlabs/genesys-ava-skills

Hi Jason,

thanks for your reply. An automated testing tool is on my Christmas wish list now. Maybe Genesys will provide something like that in the future. The integrated Testing tool from parloa is impressive.

Best regards
Christoph

Christoph,

 We are facing a similar issue, we had been conditionally rolled back to a previous opus on our org due to unexpected issues when Genesys updated the haiku model they were using to support freeform slots. I anticipate the v2 agentic stuff is in a similar boat. We've had to kind of scramble to re-test, but it appears it's less something Genesys has done and more with v3 being retired. 

Claude 3 Haiku - Amazon Bedrock

Best of luck with the testing! Hopefully you can employ an automated testing tool to decrease the lift.

The SDD skills are truly a game changer for building, as well as regression testing. You will very easily be able to set up a robust evaluation suite, choose the new model and re-run it. Any failures are generally also easily fixed by Claude et al when using the SDD. 

We also have roadmap items (announced this week at our Xperience conference) for pulling all of the SDD skills directly into the GC Copilot, so it will all live right next the UI as you build. 

In our benchmarks, very little changes were necessary for apt1->2., which is different than Jason's previous post of haiku 3->4.5 and was done for different reasons. 

Haiku 3 was deprecated by AWS, whereas we are managing the models for apt1/2 so we want to deprecate in order to ensure we have GPU capacity for newer models. This gives us a little more flexibility though we need to be aggressive. Every model change will be handled to the best of our ability. 

Hi Savino,

thanks for your insights. Good look with the change.

Thanks for the link, too. I will take a deeper look.

We had an earlier preview of the APT-2 model and from my experience we're having to do a fair bit of rework to make it behave as it did with APT-1. in terms of testing, Genesys released Spec Driven Development for AVA which may be worth a look at. 

https://github.com/purecloudlabs/genesys-ava-skills

Hi Jason,

thanks for your reply. An automated testing tool is on my Christmas wish list now. Maybe Genesys will provide something like that in the future. The integrated Testing tool from parloa is impressive.

Best regards
Christoph

Christoph,

 We are facing a similar issue, we had been conditionally rolled back to a previous opus on our org due to unexpected issues when Genesys updated the haiku model they were using to support freeform slots. I anticipate the v2 agentic stuff is in a similar boat. We've had to kind of scramble to re-test, but it appears it's less something Genesys has done and more with v3 being retired. 

Claude 3 Haiku - Amazon Bedrock

Best of luck with the testing! Hopefully you can employ an automated testing tool to decrease the lift.

Very impressed with SDD for AVA so far. Being able to build, critique, create test cases and test (with mock data actions!🤯) and publish an Agent just by having a conversation in the terminal really is game changing. 

Although we are heavy users of CX as Code, it would be great to see skills developed for other resources in Genesys Cloud.

The SDD skills are truly a game changer for building, as well as regression testing. You will very easily be able to set up a robust evaluation suite, choose the new model and re-run it. Any failures are generally also easily fixed by Claude et al when using the SDD. 

We also have roadmap items (announced this week at our Xperience conference) for pulling all of the SDD skills directly into the GC Copilot, so it will all live right next the UI as you build. 

In our benchmarks, very little changes were necessary for apt1->2., which is different than Jason's previous post of haiku 3->4.5 and was done for different reasons. 

Haiku 3 was deprecated by AWS, whereas we are managing the models for apt1/2 so we want to deprecate in order to ensure we have GPU capacity for newer models. This gives us a little more flexibility though we need to be aggressive. Every model change will be handled to the best of our ability. 

Hi Savino,

thanks for your insights. Good look with the change.

Thanks for the link, too. I will take a deeper look.

We had an earlier preview of the APT-2 model and from my experience we're having to do a fair bit of rework to make it behave as it did with APT-1. in terms of testing, Genesys released Spec Driven Development for AVA which may be worth a look at. 

https://github.com/purecloudlabs/genesys-ava-skills

Hi Jason,

thanks for your reply. An automated testing tool is on my Christmas wish list now. Maybe Genesys will provide something like that in the future. The integrated Testing tool from parloa is impressive.

Best regards
Christoph

Christoph,

 We are facing a similar issue, we had been conditionally rolled back to a previous opus on our org due to unexpected issues when Genesys updated the haiku model they were using to support freeform slots. I anticipate the v2 agentic stuff is in a similar boat. We've had to kind of scramble to re-test, but it appears it's less something Genesys has done and more with v3 being retired. 

Claude 3 Haiku - Amazon Bedrock

Best of luck with the testing! Hopefully you can employ an automated testing tool to decrease the lift.

The SDD skills are truly a game changer for building, as well as regression testing. You will very easily be able to set up a robust evaluation suite, choose the new model and re-run it. Any failures are generally also easily fixed by Claude et al when using the SDD. 

We also have roadmap items (announced this week at our Xperience conference) for pulling all of the SDD skills directly into the GC Copilot, so it will all live right next the UI as you build. 

In our benchmarks, very little changes were necessary for apt1->2., which is different than Jason's previous post of haiku 3->4.5 and was done for different reasons. 

Haiku 3 was deprecated by AWS, whereas we are managing the models for apt1/2 so we want to deprecate in order to ensure we have GPU capacity for newer models. This gives us a little more flexibility though we need to be aggressive. Every model change will be handled to the best of our ability. 

------------------------------
Mitchell
Principal Product Manager, Virtual Agent
------------------------------


Original Message:
Sent: 09-04-2026 02:24
From: Christoph Domaschke
Subject: deprecation of Agentic Virtual Agent Model Version 1

Hi Savino,

thanks for your insights. Good look with the change.

Thanks for the link, too. I will take a deeper look.

------------------------------
Christoph
Produktmanager Kunden-Dialog-Center
------------------------------



Original Message:
Sent: 09-03-2026 07:40
From: Savino Ricci
Subject: deprecation of Agentic Virtual Agent Model Version 1

We had an earlier preview of the APT-2 model and from my experience we're having to do a fair bit of rework to make it behave as it did with APT-1. in terms of testing, Genesys released Spec Driven Development for AVA which may be worth a look at. 

https://github.com/purecloudlabs/genesys-ava-skills

------------------------------
Savino Ricci
Senior Technical Consultant
------------------------------




Original Message:
Sent: 09-03-2026 03:14
From: Christoph Domaschke
Subject: deprecation of Agentic Virtual Agent Model Version 1

Hi Jason,

thanks for your reply. An automated testing tool is on my Christmas wish list now. Maybe Genesys will provide something like that in the future. The integrated Testing tool from parloa is impressive.

Best regards
Christoph

------------------------------
Christoph
Produktmanager Kunden-Dialog-Center
------------------------------





Original Message:
Sent: 09-02-2026 10:07
From: Jason Okowitz
Subject: deprecation of Agentic Virtual Agent Model Version 1

Christoph,

 We are facing a similar issue, we had been conditionally rolled back to a previous opus on our org due to unexpected issues when Genesys updated the haiku model they were using to support freeform slots. I anticipate the v2 agentic stuff is in a similar boat. We've had to kind of scramble to re-test, but it appears it's less something Genesys has done and more with v3 being retired. 

Claude 3 Haiku - Amazon Bedrock

Best of luck with the testing! Hopefully you can employ an automated testing tool to decrease the lift.

------------------------------
Jason
Enterprise Architect
------------------------------

We had an earlier preview of the APT-2 model and from my experience we're having to do a fair bit of rework to make it behave as it did with APT-1. in terms of testing, Genesys released Spec Driven Development for AVA which may be worth a look at. 

https://github.com/purecloudlabs/genesys-ava-skills

Hi Jason,

thanks for your reply. An automated testing tool is on my Christmas wish list now. Maybe Genesys will provide something like that in the future. The integrated Testing tool from parloa is impressive.

Best regards
Christoph

Christoph,

 We are facing a similar issue, we had been conditionally rolled back to a previous opus on our org due to unexpected issues when Genesys updated the haiku model they were using to support freeform slots. I anticipate the v2 agentic stuff is in a similar boat. We've had to kind of scramble to re-test, but it appears it's less something Genesys has done and more with v3 being retired. 

Claude 3 Haiku - Amazon Bedrock

Best of luck with the testing! Hopefully you can employ an automated testing tool to decrease the lift.

Christoph, I think your concern is reasonable. For teams with AVAs already in production, the effort includes testing, investigating differences, making adjustments, and validating those changes again. Thirty days can become a tight window across multiple journeys.

In our AVA projects, we've found that an appropriate response doesn't always mean the expected action or transition happened. That's why I'd want migration testing to cover tool execution, escalation, returning to a menu, and ending the conversation, alongside answer quality.

Your comment about automated testing also resonated with me. We've been exploring AI-assisted testing of AVAs, and one valuable lesson has been the importance of defining the expected outcome for each scenario. Running more conversations helps, but we still need clear criteria for deciding whether the behaviour is acceptable.

For this migration, I would use previously validated scenarios as a baseline, including variations in wording and context, and repeat them against the new model.

I'd welcome a longer validation window and clearer guidance on expected behavioural differences. For those already testing the new model, where have you needed the most adjustment: instructions, tool usage, or conversation endings?