Important Update -- Rovo Agents are getting a new backend LLM model

[#1]

UPDATE TO THE TESTING WINDOW.

We have extended the window for testing -- you have until Apr 14th to test and migrate your agents. Read more here.

---

This article is to announce an upcoming change that will begin rolling out over the next few days.

We’re upgrading the backend model that powers Rovo agents.

What to expect:

• Nothing visible. It’s a backend change only. We’re upgrading the underlying model behind the scenes.

• Configuration stays the same. Agent instructions, skills, knowledge, and permissions don’t change.

• New agents: any agents you build after March 17th (US Date), tomorrow, will use the updated backend model by default.

• Existing agents: An existing agent can be switched to the updated model throughout March. You can swap between the old model and new model to iterate on your agent and make sure it performs as expected.

After March 31st (US Date), any remaining agents still using the old model will be upgraded automatically.

You will see this is live when you notice a purple banner in Studio or in your Agent page announcing the change. 

How to test the improved model

There are two ways to test your agent on the new model

1. Duplicate your existing agent:

   1. From the agent menu, select Duplicate.

   2. The duplicated agent will automatically use the new backend model.

   3. You can run LLM evaluations or manual tests on it to validate behavior.

   4. When you’re confident in the output you can either keep the duplicated agent, or delete it and update the original agent to the new model (you have until the end of March).

2. Use the model "switch" in Studio

   1. Edit an existing agent

   2. In agent overview page, look for Swap to new model

   3. Toggle it ON

   4. Either run LLM evaluations or manually test your agent, you can go back into edit and toggle it off, if you want to go back to the old model. However, you’ll only be able to do this until the end of March when all remaining agents using the old model will be upgraded automatically and the toggle will be removed.

FAQs

• Will this change how my agent behaves?

  • For most agents, behavior should remain consistent. Because the underlying model is changing, you may see some differences in multi-turn conversations or how responses are phrased. Validate high-value agents with evaluations or key prompts before switching.

• Can I keep using the old backend model indefinitely?

  • No. Use the 14-day window to test and adjust any critical agents before we upgrade remaining agents.

• Who do I contact if something looks off?

  • Contact Support with the agent link and include “Agent model upgrade” in the description.

• Why are you changing models?

  • We’re constantly evaluating LLM models as they are released. With newer flagship models available, we’ve decided on one that gives the right balance of access to the newest models while avoiding major changes for existing agents.

• Where do I find the toggle? 
  • The toggle can be found in Agent Edit -> in the Identity tab. 

[hayato yamazaki]

Will this model change affect the side panel Robo that I use when I open a Confluence page?

Will the changes apply to the side panel Robo and the search Robo as well?

[Piotr Smialek]

Does the new model handle anything better? If you were able to test the change and could share the differences with examples, it would be easier to verify the models' performance with customers.

[Stefan Kirchner]

Hello,
Is there an overview or a list of what has been improved/changed?
Thanks.

[Colton.Kosicek]

What is this new LLM model? we are not allowed to use certain LLMs this will force companies to disable AI very quickly if you do not share more details ASAP

[Jensen Fleming]

@hayato yamazaki -- this will impact agents you speak with in Rovo Chat side bar, but not Rovo Chat without an agent. Rovo Chat uses a collection of models and is on a different rollout cycle. 

@Piotr Smialek @Stefan Kirchner  This update has shown: 

• Search / knowledge Q&A

  • Better answer rate and relevance on Search / internal policy questions.

  • More likely to pick the right search queries and tools.

• Overall answer quality

  • Higher accuracy, groundedness, and correct tool choice.

• Service “next steps” / guidance

  • Better instruction following, relevance, conciseness, and personalization.

• Email summarization / replies

  • Higher semantic quality and completeness.

• Latency & verboseness. 

  • Faster total latency when calling the same/fewer tools.

  • Shorter, clearer outputs; similar or better quality; often faster.

@Col -- all of our approved providers are listed here: https://www.atlassian.com/legal/sub-processors#atlassian-group-sub-processors. For transparency, the model used under the hood for this change is Gemini 3 Flash. 

[Stefan Hanitzsch]

Apparently Gemini-3-Flash is one of the underlying LLM. I got an error message which contained the model route. The agent said that I could not add as many skills to any one agent. I had added all of Confluence and Jira skills. 

[Jensen Fleming]

@Stefan Hanitzsch -- where did you see this error? There shouldn't be any limitation there. 

[Razvan Nechifor]

Hello, I am receiving this error on a new agent created today (23rd of March) as a copy of an existing agent (created before 17th of March):

error_category: BAD_REQUEST_ERROR, error_type: bad_request, model: gemini-3-flash-preview, status: 400, gateway_request_id: bc3ad9f2-2431901, gateway_message: , gateway_error: Bad Request, gateway_path: /v1/google/v1/publishers/google/models/gemini-3-flash-preview:streamGenerateContent, upstream_status_code: 400

The error appears any time I am trying to interact with the agent.

It seems that existing agents on which I am turning on the switch Swap to new model have the same error.

[Jensen Fleming]

@Razvan Nechifor can you open a support ticket so we can take a look? Please link it to me so I can follow up on it. 

[František Špaček _MoroSystems_]

I am experiencing the same problem on my sandbox on completely new agent. Just default one scenario, about 30 skills (almost all from Jira and Jira Service Management are added). Once I have deleted few of those, it suddenly works. Error is the same as above (contrary to the Razvan it only happens if I switch to new model).

[Induprasad S R]

Hi There, Good Day!

I was using a ROVO sub agent to get user questionaries and create an Epic based on the answers.
It was working so far, but from today when I switched to the new model, it stopped to create the Epic stating it doesnt find the respective board or don't have permission to create it.
What could have gone wrong, how should I fix this?

[u400492]

New Rovo Studio "Model" Toggle Breaks Attachment Processing (RAG)

Issue Summary: I have identified a critical regression in Rovo Studio related to the new "Model" selection feature in the Identity settings. Enabling this toggle appears to disconnect the agent from the RAG (Retrieval-Augmented Generation) pipeline.

Observed Behavior:
Model Toggle ON: The agent completely ignores uploaded attachments (PDF/Text) and produces generic hallucinations or "memory-based" summaries. It fails to "Search and Read" the provided files.
Model Toggle OFF: The agent functions perfectly, extracting and translating data from attachments as expected.

Technical Context:
Capability: "Search and read Confluence pages and files" is enabled.
Debug Logs: The logs confirm the file is received, but the content is not being injected into the new model's context window.

Urgency: The Rovo UI states a mandatory migration to this new model architecture on March 31, 2026. If this is not resolved, all agents relying on attachment analysis will break in less than a week.

Has anyone else encountered this context loss with the new models? Is there a known workaround before the March 31st deadline?

[Jensen Fleming]

@František Špaček _MoroSystems_ @u400492 @Induprasad S R -- can you all please open support tickets so we can dive into each one?

[Maurice Adams]

I wonder if this will help some of the issues I have seen between the ROVO browser ext agent I created in Atlassian that we don't see with the non browser agent.  Some times ROVO agent responds better within the suite than the browser agent. 

[sankaracourtallam_rathinasabapathy]

New Rovo model not retrieving Jira attachment metadata exhaustively

I’m using a custom Rovo agent to generate structured bug reports from Jira issues, including exhaustive retrieval of attachments and metadata. With the previous model, the agent was able to list all attachments and related evidence reliably.

After the recent model update, I’ve noticed that the agent no longer performs exhaustive retrieval. It often summarizes or omits some attachments even though the information is present in the Jira ticket. This is affecting traceability and completeness of the generated bug reports.

Is there a way to:

• continue using the previous model for an existing custom Rovo agent, or
• pin the agent to a specific model version to maintain consistent behavior?

If reverting is not possible, are there any recommended configuration or prompt-design changes to restore the earlier retrieval behavior?

[Scott Giles]

New LLM is significantly struggling when being asked to create work items, no longer listening to strict guardrails for work item type. Also providing very misleading outputs on occasion too. This change seems like it will detract from the purpose of my custom Rovo Agents 

[Terranova_ Diego Nicolas]

Hello, currently when I switch to the new model, all agent requests fail. I have tried duplicating the requests and enabling the new model, but I consistently encounter the same issue along with the following message:

error_category: BAD_REQUEST_ERROR, error_type: bad_request, model: gemini-3-flash-preview, status: 400, gateway_request_id: 7b18a1fc-1382192, gateway_message: , gateway_error: Bad Request, gateway_path: /v1/google/v1/publishers/google/models/gemini-3-flash-preview, upstream_status_code: 400

Has anyone experienced this issue or can provide any guidance?

[Jensen Fleming]

Quick update on our (Atlassian) side  -- we've extended this window to allow you to toggle between models until April 14th to give you more time to test and migrate your agents. 

I will provide a more detailed update soon but wanted to provide this update quickly.  

[Stefan Kirchner]

Hello,
In the last few days I had received the following error message with the new LLM models several times.

'error_category: BAD_REQUEST_ERROR. error_type: bad_request, model: gemini-3-flash-preview, status: 400. gateway_request_id: 1e9a4743-699630, gateway_message: Exceeded limit on max bytes to buffer: 104857600, gateway_error: Bad Request, gateway_path: /v1/google/v1/publishers/google/models/gem ini-3-flash-preview:streamGenerateContent'

Anyone else who gets this? Any ideas?

[Veranika Artamonava]

Same for me.
I have the error all the time, and I unfortunately deleted my agent on previous model :( 

error_category: BAD_REQUEST_ERROR, error_type: bad_request, model: gemini-3-flash-preview, status: 400, gateway_request_id: a69e3c1f-2447216, gateway_message: , gateway_error: Bad Request, gateway_path: /v1/google/v1/publishers/google/models/gemini-3-flash-preview:streamGenerateContent, upstream_status_code: 400

[sankaracourtallam_rathinasabapathy]

Regression: Rovo agent (New Model) no longer lists all Jira attachments

I’m currently facing an issue with the new backend model used by Rovo agents (rolled out around April 14, 2026). In my use case, I have a custom agent that formats Jira bugs and is expected to retrieve and list all files from the Attachments metadata.

With the previous model, the agent was able to enumerate all attachments reliably even with relatively minimal instructions. However, with the new model, it is no longer able to retrieve files that are present in the primary Attachments field. It only seems to access attachments that are referenced through comments or activity, which leads to incomplete evidence listing.

I’ve spent a significant amount of time experimenting with prompt refinements throughout the day — including:

• enforcing strict enumeration rules
• adding validation steps
• specifying retrieval from both Attachments and Activity metadata
• prohibiting summarization or grouping

Despite these refinements, the new model still omits several attachments or replaces them with generalized entries, even when the files are clearly present in the ticket metadata.

In contrast, the older model handled this scenario much more reliably and produced complete attachment listings with far simpler instructions.

Is this a known regression or a change in how attachment metadata is exposed to the model?

Has anyone else observed reduced attachment metadata retrieval with the new model?
Is this a known limitation or change in how attachment data is exposed to agents? Any guidance or recommended workaround would be appreciated.

[Colton.Kosicek]

My company is becoming increasingly concerned that we are being limited to FLASH and not allowed access to thinking or PRO models. this will greatly restrict our output capabilities.

[Justin Townsend]

Constructively, it might help customers / solutions partners / consultants if you were to consider publishing your reasoning (e.g. Model Scores or other) for the change?

An LLM shift can be disruptive for even modern processes built on earlier LLM versions.

Understand AI integration is crucial for business, just gotta make sure clients are all ready adopt it with the volatility it brings to established prompt scores.

:-)

[Neha Asthana]

Has anything happened to resolve this error? We are building serious applications using rovo chat agents and this will halt many things for us for further development. Can someone please look at it urgently?

error_category: BAD_REQUEST_ERROR, error_type: bad_request, model: gemini-3-flash-preview, status: 400, gateway_request_id: **, gateway_message: , gateway_error: Bad Request, gateway_path: /v1/google/v1/publishers/google/models/gemini-3-flash-preview:streamGenerateContent, upstream_status_code: 400

[Jensen Fleming]

@Neha Asthana @Veranika Artamonava  @Terranova_ Diego Nicolas -- We have rolled out a fix for this issue, please let me know if it is now working for you. If not, can you please open a support ticket so we can look at the logs. 

@Stefan Kirchner @Razvan Nechifor -- Did you open a support ticket for this? Can you link it to me please so we can look at the logs, this is different error than the one above. 

@sankaracourtallam_rathinasabapathy -- We are working through attachment issue, will report back shortly. 


@Colton.Kosicek - totally hear you, we are exploring Model Category/Class selection, and hope to provide that soon -- https://jira.atlassian.com/browse/ROVO-275. Will provide an update on that ticket this week.