Atlassian on AI Agents, Teamwork Graph, and the Future of Work

[00:00:00] Do you need AI Agents that you can trust? Well, with an AI Data Layer providing real-time connection within your data platforms, you can trust your Agents to provide accurate solutions. So, scale your business by trusting your agentic AI accurately getting the work done for you. Trust its capabilities with Denodo. And you can do that by simply visiting denodo.com to learn more.

[00:00:26] What happens when AI stops behaving like a chatbot and starts behaving like a teammate? This is a question that sits at the centre of everything happening here at Team 26 in Anaheim. Because over the last few years, we've all watched AI evolve from simple assistants helping us to summarise documents or generate snippets of code into something far more ambitious and exciting.

[00:00:53] Because now the conversation is shifted toward agents, orchestration, organisational context, and AI systems capable of carrying work forward alongside their human teammates. But yeah, there is a problem still here. Many AI tools are still operating in isolation. They don't understand that relationship between people, projects, documentation, decisions, or workflows inside a company.

[00:01:20] And my guest today will tell you that missing context may be the biggest thing holding enterprise AI back right now. And my guest today leads engineering across products that you all know and love from Jira, Confluence, Loom, Trello, and all things AI at Atlassian. Including Rovo and the rapidly expanding teamwork graph. So today we will unpack why contact is becoming such a strategic advantage in enterprise AI.

[00:01:45] Why Atlassian believes that open ecosystems matter far more than closed AI stacks. And we'll discuss how the company is trying to transform AI from just isolated assistants into something that is deeply embedded inside the flow of work itself. And we'll also talk about the rise of AI native engineering teams, smaller autonomous squads, agents taking ownership of operational work.

[00:02:09] And how Atlassian has managed to build products like Confluence, Remix, and Teamwork Graph in just a matter of months. The timeline on that is breathtaking, let me tell you. So if you're wondering where AI is really heading after the chatbot phase, today's conversation should give you a fascinating look at what might come next, the ROI, and the speed of which you can bring all this to life. But enough from me. Let me introduce you to my guest now.

[00:02:37] So thank you for joining me here at Team 26. For everybody listening, hearing about you for the first time, can you tell them a little about who you are and your role here at Atlassian? Well, thanks for having me on this podcast. My name is Tarun. I'm the CTO of Atlassian. I lead some of the products that you might have used. Jira, Confluence, Loom, Trello. So responsible for engineering for those products. I also lead all our core engineering teams. And then lastly, all things AI.

[00:03:03] So if you've heard of Rovo or any of our AI capabilities, I'm responsible for the engineering for that. And it really feels like an exciting time for you at the moment. And one of the big themes coming out of Team 26 is AI. Yes, it is everywhere, but it should behave more like a teammate than a chatbot. So what does that actually mean in practice? So why do you think so many AI tools today still feel disconnected with how real teams work? Because there's a lot of examples of that and a problem that you're solving here. Absolutely.

[00:03:31] So I do think obviously the role that AI tools are playing is changing every passing week because of the capabilities of the models, which are getting really, really good in terms of reasoning. Also, the skills that we are giving to these models in terms of what tools they can call. So definitely, they are getting better every passing week, yet they feel disconnected to your point. And part of the reason is you really want to give them the context that a human has working in the team.

[00:03:56] If you had an intern or even a junior engineer who's on the team, they often would understand a lot about who are the people, who to go ask for, where is the information, what has been done before, where does the code belong. A lot of that context is very useful for humans. And guess what? That context is also very useful for agents because grounded in that context, they can then go produce outcomes or they can produce output that is really relevant to what you're trying to solve.

[00:04:24] So one part of really making AI work within a company is making sure you are arming it with the right context. And I think the ones who are able to nail that better are probably leading to better outcomes. And to be honest, that's exactly the problem we are also going after. To say we, our company has carried, I guess, work context for years. We have more than 20 years of what, who, why and off work. And if we can start to bring that in a fashion that agents can consume in a very effective

[00:04:53] way, hopefully we can improve their output. We can make it cheaper for them to, to produce that output and hopefully faster. And that word there, context, it's a big, big theme here. And for the techies listening as well, you're opening up the teamwork graph through MCP and CLI access, effectively allowing organizations to bring their operational context into any AI tool. So why is context becoming such a strategic advantage in enterprise AI?

[00:05:21] And what problems do you see it solving that maybe generic AI systems can't? Yeah, absolutely. Like at the end of the day, I mean, these are extremely powerful models. If you really arm them with what, who, why, they can produce far relevant output to the problem you're trying to solve. I'll take a few examples, right? When you're using a coding agent, most of the coding agents, people might be using them in an IDE or they might be using them in a CLI.

[00:05:50] What is accessible to them is the code in front of them. So of course, they're able to span the repo. They can scan it. They can try to reason out on that repo. You can give them some context from a Jira ticket and say, here is what I'm trying to solve. Here is the code. Go help me figure this out. But oftentimes for a human, if you think about it, a lot of maybe the runbook around that repo may be sitting in a Confluence page. Maybe the PRD for that ticket that you're trying to solve for is perhaps sitting in a Google doc somewhere.

[00:06:18] Maybe there's some additional test case context that might be sitting in Slack. There might be prior work done in related issues which might be sitting in code or other Jira issues. So if your AI is able to get access to some of that, that can lead to a far more relevant output that doesn't miss certain edge cases, that doesn't miss certain conditions or what you're trying to do. So the more and more you are able to bring that, I think it'll improve the output of that AI. And I'm giving just a coding example.

[00:06:46] The thing applies for many, many other scenarios as well. Let's say we are planning a team event like a sales marketing team is actually trying to set up an event like the one we are at right now. I think, again, a lot of context on what happened in prior teams, what things worked, what things didn't work. How should you actually organize your team sessions? A lot of that history is sitting in Google Sheets and it's sitting in Confluence pages and a bunch of Jira boards from our previous team.

[00:07:14] Now, imagine if you're using AI to go plan out this current team event. A lot of that history is going to be super valuable to understand what worked and what didn't work so that they can produce a much better plan for the next one. So I do think we are very convicted that this context is, and every enterprise has a pretty unique context grounded into the problems they are going after, grounded into the customers they are serving. So if we are able to bring that in a very efficient way to the agents, they can do a better job.

[00:07:44] Now, I would also add that a lot of the agents, like you might be using Cloud Cowork, you might be using ChatGPT, they do allow you to connect to some of these tools. So of course that helps. But then again, a lot of this information, first of all, is a lot like they allow you to connect to some may not allow you to connect to all. Secondly, I think there's a lot of derived information in terms of relationships, which it takes a while to go process it and understand those relationships.

[00:08:11] And that's the difference between a correct output or incorrect output and something of that nature. We try to build it ahead of time. And largely, sometimes you might be thinking of insane amount of data. So many of these tools can go look up a Jira issue quickly. They can go look up a Confluence page. They can go look up a Figma design. But being able to search through your entire company's context, particularly if you're thinking about companies who have been around for a while, have a lot of context, you really

[00:08:37] need a very efficient way to be able to understand those relationships or to be able to scour through that data. And that's where search comes into play. So we have been building this for a while. And we are already like we were able to see a lot of improvements with our own agent robo. And then we realized we should actually make it available to other agents as well, because people want to work and use the agents they want to use. If we are able to bring that through MCP or CLI, hopefully we can help those users in terms of improving the output of the agents they're using.

[00:09:06] Here at Team26, Atlassian is stating that teamwork graph now maps more than 154 billion connections across people, projects, codes, docs, and goals. At this kind of scale, how do you stop AI from becoming just overwhelming noise instead of making it genuinely useful and actionable for teams? Because it's easy to feel overwhelmed with that kind of... Yeah. So that's the whole point in a way that you want comprehensiveness because you don't want

[00:09:32] to miss a piece of context that's super relevant for the task you're looking for. So that's why that number being 150 billion is more talking to the fact that we... The graph is comprehensive. But that doesn't mean that you actually give all of that data to AI. And that's where building a very powerful search on top comes. That's where building a very powerful graph comes because you're actually sussing out the right relationships. And then there is a very efficient way to go back and get that information.

[00:10:02] I would say it's... I agree with what you're saying. And therefore, we have built search on top so that you're not really overwhelming AI. You are still giving a very efficient way for them to be able to go access it. I will also say for that to happen, you need very high quality search. Search is a hard problem. I've worked on search for 15 years. Even before coming to Atlassian, I led the search team at Facebook. I also led the Bing search team, various parts of that team.

[00:10:28] And on Atlassian, we have a lot of talent which has experience in search. It's one thing to build search. It's another thing to build very high quality search, which is relevant. So we have made that investment. So that's why I think that's a big differentiator. I would also say semantic search has gotten much better over time. And that also helps because oftentimes you may describe things in one way, but they might be in your content in another way. So being able to semantically match is the second part.

[00:10:56] And the third part is when we are able to suss out these relationships, then giving a very efficient way to go through a pretty big graph, but really very quickly execute multi-hop queries. And we have particularly optimized our graph query interface as well. So with all these three things, typically when you connect Teamwork Graph through MCP or CLI to one of these agents, you would notice if you were able to look at the trajectories,

[00:11:22] you will actually go and notice that they are making far fewer calls because they are being able to point and shoot in a very, I would say not a predictable manner, but they are able to actually in a very effective manner, get back the information they need. And if you did not have this, they'll typically be scanning a lot more. And if they scan a lot more, their context windows blow up. The output quality gets worse at some point in time because of the window bloating. And of course, you're spending more tokens.

[00:11:48] So tell me a little bit more about the journey that you've been on and how it was built. Yeah. So I think, I mean, much like we are building solutions that other teams can use to become more productive, to get more out of, get more done. I think we have been also embracing new ways of working or AI native ways of working within our engineering team. So I'm particularly proud of how our teams have changed the way they work and embrace AI or leverage AI in the entire SDLC.

[00:12:16] I mean, some of the announcements that you saw today, like many of these things we have built, like I'll give you an example. We built the whole TWG CLI from idea to open beta today in two months. Confluence slides, if I recall, was built in eight weeks and slides is a pretty complicated content type. We built the remix and confluence in six weeks and we are starting to see these examples everywhere. But that also required us to fundamentally change how we are working.

[00:12:45] So one, of course, we are leveraging AI in many parts from planning to orchestrating work with human and agents to even on the right of code in terms of deployments and then managing the code in production. But more importantly, how the teams are also working with each other is changing. So for instance, we are actually having squads which are much smaller in size. Oftentimes, like I'll take the Confluence remix as an example, it's a combination of a PM,

[00:13:11] a designer, one or two engineers, and a very, very tight loop between them as they are going and iterating on a particular feature, which is quite different than previously where you'll throw a lot more people and you'll operate in a somewhat of a, I won't say a waterfall fashion, but somewhat of a structural fashion, bringing these people together and then really doing a very fast loop. It becomes pretty critical because AI is able to really accelerate a lot of code gen, but then the bottleneck shifts in terms of alignment and in terms of making those decisions.

[00:13:39] So that's where having a very tight loop with a few engineers and a PM and a design is really speeding things up. I would also say that roles are getting fuzzier. I think our engineers are taking a much stronger product engineering mindset. They are not just limited to code. They're very much grounded into why are we building this? What is the outcome we are making, what we are driving towards, which means then they have the agency to make decisions. If they have to make a product decision, they don't need to go back to a PM.

[00:14:08] They can make a lot of decisions because they align on principles ahead of time. Engineers are thinking about taste. Similarly, our PMs are also checking in code for some of these projects. Of course, not big check-ins, but at least they're able to go do a lot of fit and finish type of things. The ideas are starting with building a live prototype first so that you can hash out a lot of details and get it to a very high fidelity intent. So a lot of that ways of working, especially with AI assisting all of it is really helping

[00:14:36] us to speed our own development in these smaller squads. And one part is you see a lot of vibe coded sort of prototypes here and there or pieces for us. It's different. We have existing products and then we are adding these capabilities on top of those existing products. So being able to move very fast in a brownfield sort of environment has been pretty interesting and learning experiences for us. We're very excited about what we are seeing in pockets and now we are looking at how do we scale this to our entire engineering team.

[00:15:04] Yeah, the pace there of that time around is just phenomenal. And another thing that stood out in the announcements was the idea of AI agents actually owning work inside Jira. So how far are we now from AI moving beyond just assistance into genuine operational responsibility and where do humans still need to remain firmly in control? Again, a bit of a balancing act. Yeah, yeah. It's a great question. Honestly, anything which is high stakes, humans will always be in the loop.

[00:15:33] Anywhere where you're making high strategy or high leverage decisions, humans will be there. So what to build is a squarely human decision. Tradeoffs is a human decision. I would even say decision to ship something or not ship something is a human decision. Even for now, we actually still require human review after like any action of the AI. Now over time, there are certain parts we may realize where AI is getting really, really

[00:16:01] good and we may automate it. We have already started to do a little bit of that in our engineering teams. For instance, we realize that a lot of vulnerabilities, AI is able to simple ones, they're able to fix so that they can keep our code bases in a healthy shape. We have really automated all of it. We still require a human to finally review the code and merge it. But that's an example where like AI is able to automate till 90%, but we still have a human input there in the end. At some point, you could imagine we might start to do auto merge there.

[00:16:30] We do a lot of experimentation cleanup and that's all done by AI right now. That's something we have automated. So I would say probably low stake tasks where you're very strong hardening and there's a lot more confidence. And even if they make a mistake, the consequences of that is very small. I think are the places where teams might get comfortable in terms of slowly automating that. But I would say everything else, I still expect humans to be there. Although the role they play might go a little higher and AI does 80, 90% of the work and you go in and check the final output.

[00:17:00] Yeah, makes sense. And you're also supporting third party agents from companies like Figma, GitHub Copilot, Databricks and so many others. And you seem to be doing this rather than forcing customers into a closed ecosystem. Was that openness a strategic decision from day one? And how important will interoperability be in the next phase of enterprise? Yeah. So Atlassian has always been like believes an open tool chain.

[00:17:26] In fact, even much before AI, I don't know if we used to talk about open tool chain where our tools very freely connect to all these other tools. There are no restrictions. Jira, the number one source code management tool it's connected to is GitHub, even though we have Bitbucket as one of our products. So I think that philosophy or that principle doesn't change. And that's the reason we also want to. And Jira is where prior to AI, a lot of work of humans was getting planned and orchestrated.

[00:17:51] Now that a lot of that work is being done by agents, we are opening that up so that any agent can pick up that work. It's fairly open where it's not restricted to our own agent robo. In fact, our hope is that every agent out there that people want to use, they have ability to go hook it up in Jira in standard ways. So we'll keep it open. The other thing is also how the agents interact and provide traceability back to Jira so that people can still continue to use Jira as the system of record in terms of auditing agents

[00:18:20] work, understanding what they did. All of that we are standardizing such that all these agents can bring it back to Jira. So I mean, if I understand your question, like we are very much committed to making it open. That's the reason also where we are taking our context. We debated internally and we're like, no, we're going to make it available to agents outside because at the end of the day, being, taking this open approach has served us well. We think that's the right thing to do. Yeah, completely agree with you there.

[00:18:47] And there's also the feeling that the boundary between human work and AI work is starting to disappear slowly. So from your perspective, what does that healthiest version of human and AI collaboration, what does that look like in a typical organization or maybe even your own? Yeah. So look, we are all figuring this out. At this point, it's starting with simple things that AI can do. But as AI is getting powerful, there is more and more things AI is able to do.

[00:19:13] So I also think that there are places where even if it is very high stakes work, AI is still able to help human in terms of doing a better job in doing that work. So I think every passing week, I get amazed by, oh yeah, this thing becomes much easier with AI. Simple sort of tasks. But now AI is able to do somewhat more complicated tasks relatively well.

[00:19:42] But we are starting to see a lot of use of AI in terms of planning, where you go from some high level requirement to creating a very high fidelity technical spec. This is particularly grounded in your code, grounded in all your enterprise knowledge and then producing a plan and then continuing to iterate a human, iterating back with the agent to further refine the plan to a level where it feels like, okay, this is a really good technical spec. Now I can take it and now I can hand it off to agents.

[00:20:09] So I do think that's an example of something where human and agents are together working on a plan, not something that we would have imagined a year ago. I also see the same thing on the write of code in terms of responding to alerts, where an AI is able to go and help understand why this alert fired. And a human can leverage that AI to debug things. So we are seeing AI in multiple places where AI and humans are sort of working together with AI assisting. And at some point, some of those activities might become fully automated for the simpler tasks.

[00:20:39] So I think it's an evolving spectrum of where it goes. There also seems to be a big focus on the announcements between the lines of turning static documentation into something more visual and consumable. And this is such a massive thing for anyone that's worked in an organization, just trying to find version one, version three, version final and all that things. Are we entering a phase where AI changes, not just how work gets done, but how that information itself is actually communicated inside?

[00:21:09] Oh, absolutely. I actually do think being an engineer all my life, I asked every time somebody asked me a question, the real source of truth is code, especially if you're answering a product question. And oftentimes the documents trail that. And there is an extra work to somebody's at some point goes and updates the API, updates the document to bring it in line with code. Now imagine if that particular thing is fully automated, where anytime you make a code change,

[00:21:36] and this is something we are working on, your documentation is auto updated. And that's where I and your memories are auto updated. So I do think we are at this phase where AI can help keeping some of that knowledge in sync with the truth in code, truth and product. So that's, that's super powerful. I see that playing and that goes beyond code. I would also say like, oftentimes when we look at project status, and I'm talking our world,

[00:22:02] which is because JIRA is all about work or Atlassian is all about helping teams work better. Oftentimes, we use our own projects and humans come in, come collect a lot of status from multiple people, then go back and communicate to the rest of the stakeholders. And that process happens on particular rhythm that they might be working on. You could imagine that some of these things might become such that they're always up to date.

[00:22:28] Because we have the lowest unit of work is typically a JIRA work item, or a PR, or some lowest unit that our systems are connected to. So imagine if we are able to aggregate, synthesize, and have that up-to-date status in our products. So anybody wanting to know where is this focus area, where is this particular project, where is this initiative, what are the risks, you're always getting the live information. So AI does help us, especially once we are able to glue all of this together, then Rovo

[00:22:56] or many other agents that you might be using with our teamwork graph should be able to give the most recent version or the state of the system. And what that does is it actually increases the speed of the work. Because the time between, like if you have more accurate information at your hands, you can make decisions faster, you can make more accurate decisions. And that makes your whole clock speed go faster. Yeah, yeah. And we will have many people listening inside organizations that are still struggling to

[00:23:24] move beyond those isolated AI experiments. So from what you're seeing across all your customers, what separates those organizations that are genuinely operationalizing AI successfully from those that are still stuck in pilot mode? What trends do you see there? Yeah, I do think it's many things. Like this one was initially a thing. Now I see this across the board. First, the leadership and the team has to be bought in that we really want to lean in, which we are seeing across the organization.

[00:23:52] Second, I think they should be ready to transform themselves because it's definitely changing the way teams are working. I think putting some energy in some sort of an enablement team inside a company definitely helps. Because, you know, the people who just give a bunch of tools to the companies and say, go figure to the employees and say, go figure it out. They're not able to go that far. People who are able to actually not only give the tools, but also give space to experiment and then have some sort of a team that is there to do enablement.

[00:24:22] If something is working on some aspect of work, how do you go amplify and then scale it out to the rest of the organizations? Folks who are actually spending energy in enablement and in terms of, I would say, amplification of the right use cases within the company, they are seeing more success. Number two, I would also go back and say, if the more they can be grounded back into their own work to understand where and how the work flows within the company, where are the friction points?

[00:24:51] And then really getting grounded in their own work and figuring out these are the friction points where I would like to go experiment with AI to see if I can simplify this thing. Can I produce better outcomes? Can I take something that takes days and do it in minutes because I'm able to do that? Like, so folks who are leaning in with their own context are seeing more success. And at the end of the day, I would say it does require almost like a product mindset to AI deployment.

[00:25:18] And with product mindset is you need to have a very strong experimentation and iterative sort of approach. You try things, some will work, some will not work. The ones that work, you want to pour more gas on it. The ones that don't work, you want to quickly fail and then go try something else. So that those organizations, I do see them moving much faster. Love it. And finally, if we were to put all the announcements, everything that you've heard, all the conversations you've had, everything you've seen on stage, et cetera, put all of that into an LLM.

[00:25:47] What excites you most about everything when you put it all together? I think I'm most excited about, okay, there's too many favorites there for me because my teams worked on it. I was involved in some of them, but I honestly feel bringing code intelligence gives a step function sort of a dimension to our graph that just makes it extremely powerful. Being an engineer, as I said, eventually you can trace back everything to a line of code.

[00:26:16] That's where the ground truth is. So if you are able to bring that code intelligence into context and then give it to AI, particularly for software teams, that's a very massive unlock. Because anytime you ask a product question, if it's grounded in code, you'll actually be able to answer it as accurately as possible. It opens up a lot of possibilities across the entire SDLC in terms of how these agents can help in every aspect.

[00:26:43] Being an engineer, being an engineering leader, to me, I feel very excited about that unlock and what it will do to how software teams work and how they produce. I think that is a thought-provoking moment to end on. So I will add links to everything that we've discussed today so people can dig a little bit deeper. I'll also put a link to your LinkedIn as well if anyone would like to connect with you. But just thank you for sitting down with me today. I appreciate your time. Absolutely. Thank you for having me on. One of the many things that stood out to me in today's conversation was this idea that

[00:27:11] AI without context is ultimately limited. We spend so much time talking about model performance, reasoning and token windows, but the bigger challenge inside organisations might actually be coherence. Understanding who owns what, where knowledge lives, how teams operate and how decisions connect across the entire business. That's really the thread running through everything that Atlassian has announced here at Team 26.

[00:27:38] Whether it's teamwork graph, rovo agents, AI native engineering workflows or simply the push towards live operational intelligence, the bigger ambition seems to be creating systems where AI can participate meaningfully inside the flow of work rather than just operating as an isolated assistant on the sidelines. And as always, I'd love to hear your thoughts on this episode. Are we moving towards a future where AI agents genuinely become operational teammates inside organisations? Are we already there?

[00:28:06] Or do you think they're still underestimating the complexity of trust, governance and human insight? And perhaps most importantly, do you and your organisation actually have that context infrastructure that is needed for AI to succeed at scale? Let me know your thoughts as always. It's techtalksnetwork.com. 4,000 interviews there. Lots of ways you can contact and work with me. And I'll await your message. But that's it today. We're out of time. But I'll be back in your podcast feeds bright and early tomorrow.

[00:28:35] Thanks for listening. Bye for now.