Rovo vs. Dia Browser. HOW IS THIS EVEN POSSIBLE???

[#1]

HEEEELLLOOO everyone!

Look, I’m a Rovo early adopter. And I want it to work. I’ve been playing with Atlassian Intelligence since day one, and I genuinely think it’s useful for a lot of things — summarizing tickets, answering quick questions about your backlog, doing simple lookups. That stuff works great.

But after spending way too many hours trying to train Rovo agents for complex, real-world admin tasks, we need to talk about something that’s been driving me absolutely crazy. Because, do you know who does every Rovo task an incredible lot better? Dia Browser!!!!  It does what my carefully crafted Rovo agents can’t, and it does it on the first try. No training. No prompt engineering. (well, well, a bit)

No “let me try rephrasing that for the fifth time.”

Let's see:

---

1. The Assets Object Search Disaster

Here’s the scenario. I needed an agent that could search for objects in Assets (formerly Insight). Sounds reasonable, right? You’d think an AI built into the Atlassian ecosystem would have a natural advantage here — it lives inside the product, it has access to the schema, the object types, the attributes.

What actually happened: Rovo hallucinated. Spectacularly. I’m talking made-up object names, invented attribute values, confident references to things that simply do not exist in my Assets schema. I tried refining the agent instructions. I tried being more specific. I tried giving it examples. It kept confidently returning garbage.

Dia Browser?: I asked the same question in Dia. It navigated to my Assets instance, found the actual objects, read the real schema, and gave me a detailed, accurate breakdown — object type, attributes, values, relationships. First try. No training. No prompt. Just… the right answer.

The difference is brutal. One tool is guessing based on patterns. The other one is actually looking at your data.

---

2. The JSM Service Desk Field Explorer

This one really got me. I was trying to build a Rovo agent that could help me explore the fields available in a JSM service desk — request types, their fields, what values they accept, how the help center is structured. This is the kind of thing you need when you’re setting up a new service project or onboarding someone into your configuration.

What Rovo did: It gave me JQL-based answers. That’s it. “Here’s how you search for issues with field X.” Cool, but that’s not what I asked. I wanted to understand the structure of my help center. The request types, the forms, the fields within each one, the allowed values. Rovo doesn’t seem to understand the difference between “search for tickets” and “show me how this service desk is built.”

Dia Browser?: It went to my JSM portal, navigated through the help center configuration, and — I kid you not — drew me a diagram. A visual map of the help center structure showing each category, each request type underneath it, the fields inside each request type, their types, their default values, whether they’re required or optional. Like, WHAT?!

I stared at my screen for a solid 30 seconds. I had spent HOURS trying to get Rovo to give me even a flat list of fields. And Dia just… mapped the whole thing out. Visually.

---

3. Why This Is So Frustrating

Here’s the thing that makes this truly maddening: I WANT Rovo to be good at this. We are invested in the Atlassian ecosystem. We build automations, we configure workflows, we manage Assets schemas, we set up service desks. Having an AI assistant that lives inside those tools and actually understands them would be amazing.

But right now, when it comes to complex agent tasks, the experience is:

• You spend hours crafting agent instructions

• You test, it fails, you tweak, it fails differently

• You add examples, constraints, guardrails

• It still hallucinates or gives you surface-level JQL answers

• You give up and ask your browser

And this browser just goes and reads the actual pages based on your permissions, understands the actual structure, and gives you the actual answer.

The browser sees what’s really there. The agent guesses what might be there.

---

4. What I Think Is Going On

Not really sure! It seems that Dia has better access to Teamwork Graph and APIs, so it understands better the relations between users, objects, etc.

And Rovo seems so dumb sometimes next to it ¯\_ (ツ)_/¯ 

---

So… What Now?

I’m not saying “don’t use Rovo.” NONONONO
For quick ticket interactions, it’s great. For chat-based Q&A about your projects, it works.

But if you’re spending hours trying to train a Rovo agent to do something complex — searching Assets, mapping service desk structures, understanding your configuration in depth — try Dia!.

You might save yourself a whole afternoon of frustration.

And Atlassian, if you’re reading this: please give Rovo agents deeper access to real configuration data. Let them actually browse the product the way a human (or a smart browser) would. The potential is there. The intelligence is there. The access isn’t.

Would love to hear your experiences — has anyone managed to get Rovo agents to do complex schema or configuration tasks reliably? Or are you running into the same wall?

Drop a comment below. Let’s figure this out together! 👇🏻️

[janus_lorenzen_eficode_com]

WTF!! I am on the sales side of Atlassian. This is crazy. I don´t even know DIA browser...Thanks for sharing!

[Prashanth]

Hi @Juan Carlos Pin 

This is a really useful writeup, and I don't think you're wrong about what you experienced. But I think the root cause is slightly different from what you've concluded, and it's worth checking before you write Rovo off for this category of task.

What you're describing -confident, made-up object names and attribute values - is the classic signature of an agent reasoning from pattern memory instead of retrieving ground truth. That happens when the agent doesn't have the right tool enabled to actually go look at your data, so it fills the gap with something plausible-sounding instead. That's not a Rovo ceiling, it's a missing skill in your specific agent's configuration.

Worth checking directly: Assets is now part of Rovo's unified search, and there's a dedicated "Search Assets data" skill you can explicitly enable on an agent in Rovo Studio. If that skill wasn't switched on for the agent you built, it had no path to your actual schema and was always going to guess. Same root cause, very different fix than "switch tools."

For the JSM field/structure mapping problem, that's a slightly different gap, that sounds like it needs an explicit action that calls the JSM configuration API directly (request types, fields, allowed values), rather than relying on the general-purpose search skill, which is tuned for finding tickets, not introspecting schema. If no one's built that action yet for your instance, that's the actual blocker, not Rovo's intelligence.

On the memory point buried in your post - the "it forgets what it just found" frustration, that's real and well documented elsewhere in the community too. That's a context window limitation, not a hallucination problem, and it's worth treating as a separate issue with a separate fix (shorter, more scoped queries rather than one long exploratory conversation).

Last thing: Rovo picked up MCP support recently, which is squarely aimed at the gap you're describing, letting agents reach into third-party and internal systems for live structural data instead of relying purely on indexed search. If Assets/JSM schema introspection isn't fully solved by the native skill, an MCP-based connector to those APIs is probably the more durable fix than prompt engineering your way around it.

None of this is to wave away what you ran into. It clearly cost you real time. But "the intelligence is there, the access isn't" is closer to the actual diagnosis than "Rovo can't do this" , and the access gap looks fixable on your end, not just Atlassian's.

[Juan Carlos Pin]

Thx @Prashanth

My agent has indeed the proper skill assigned, but even though, it's not getting to the right answer. I'm still on Rovo train, nonetheless 😏

[Prashanth]

We all are in the same ship mate, just different rooms :) @Juan Carlos Pin 

[Jan Voigt]

Hey @Juan Carlos Pin, interesting article! I actually tried to reproduce your Assets scenario right now with Dia — and in my experience, it hit the exact same wall as Rovo.

I asked Dia to look up who is linked to a specific object in my Assets schema. It tried everything: the Teamwork Graph, multiple REST API paths, fetching the rendered UI pages — all either returned empty results, 401/404 errors, or “Unsupported product: cmdb.”

The only way I got the answer was by opening the Assets page myself in the browser and letting Dia read the page content from there. At that point it correctly identified the 3 linked people — but that’s not really “navigating to Assets and reading the schema” autonomously. That’s me doing the navigation and Dia reading what’s on screen.

So at least for Assets/CMDB, Dia seems to have the same limitation: no direct API access to the underlying data. Would love to know more about your setup — maybe there’s a configuration difference that made it work for you?

[soujanya_palem]

Thank you posting this @Juan Carlos Pin . I ran into the same issue of Rovo making up answers when trying to read Assets. 

[Juan Carlos Pin]

Hi @Jan Voigt maybe you need to grant external application permits to the assets schema, when I asked, even in a new chat to Dia, it went to my schema got the object, its attributes, everything! 

I guess that, since The Browser Company was acquired by Atlassian last year, they both will find a cool way to turn this around!

[Michael_C_Weihrauch]

Your frustrations with spending hours trying to get Rovo agents to repeat a learned task consistently—only to end up with disappointing results—sound incredibly familiar. Thank you for sharing these detailed observations. Hopefully, Atlassian is listening closely to this feedback.

[Tomislav Tobijas]

I've seen a couple of cool things around Dia, but you only need a Mac to try it out... 🫠

[Juan Carlos Pin]

@Tomislav Tobijas  I know, meanwhile → https://www.diabrowser.com/windows

OR!!! convince (wink, wink) someone to get you a mac (for all these years, for me, it's worth every penny)

[Bill Goetz]

I've used Arc/Dia in the past and it definitely has it's advantages over other browsers. The bigger issue is that Dia isn't going to work at the corporate level. Most companies are going to use a Chromium based browser over Dia.

Can only see this use case be used in specific scenarios.

[Juan Carlos Pin]

The funny and strange thing is that when tapping into the atlas’s Ian skills to perform tasks, Día shows the Rovo icons spinning when thinking. So, below, is there any part of Rovo there?

[Devlend Maul]

A key point regarding Rovo is its heavy reliance on the Teamwork Graph. To my knowledge, that doesn't yet natively include access to backend system configuration structures like Jira schemes, workflows, or permissions. Dia on the other hand operates within the actual workflows while still leveraging the Teamwork Graph, which grants it access to deeper layers of data. As with most tools, each has its own strengths and weaknesses, and it is up to us to determine which is the best fit based on our users specific needs and the overall health of our ecosystem.

[Patricia Francezi_iDev_]

i do know the feeling......... You never think  that rovo would be dumb at atlassian stuff... never ever