Atlassian CEO: Why Agents Won't Replace Interfaces | Why Enterprises Won't Trust Chinese Models

If you're giving up on your user interface, if you're giving up on evolving your experience for a customer, that's a brainless bet. I don't think agents will replace all the interfaces.

For the first time in a long time, we're solving foundational design challenges. Again, there is certainly some customer pushback in some places to quote-unquote Chinese models. There is not customer pushback to open models. The jury is really out on the benefits of building specialized models at the moment. Right now, the production of technology is faster than the consumption of technology. We're moving from a world of information asymmetry to imagination asymmetry.

Oh, thanks for having me back, man. It's appreciated. Get invited back. I must give those interesting answers. I saw your little rant the other day. I loved it. You give honest answers, which I don't always get.

I think there's some premise of the question that's slightly incorrect. I think, look, I can see for certain tasks that is the case, where the tasks are very functional, that is the case. Where the tasks have a lot of taste and interaction and collaboration involved, I don't think that will be the case, certainly not in the same way. So it's very hard to give an answer across all types of tasks. You know, one might ask, there are certain types of shopping, if you take a consumer task, not something that we do obviously. So, a safe example, there are certain types of shopping. I just want to get the thing and move on. If I'm buying, I don't know, new Nike trainers, a charging cable off Amazon, just give me the charging cable. I don't care. I don't need anything, right? And if there are other types of shopping where I may enjoy the experience, I may want to go to the store, I want to do a lot of research, I want to do it myself. There's going to be a whole range of things. And so, yes, for some tasks, it's going to get rid of the pedestrian, 'I just want to get the thing done.' But in the enterprise collaboration world and in the business workflow world, agents are going to help and are already helping, I would argue, in a lot of cases, move the work forward to let people honestly focus on higher value, creative parts of the job. I think that will change a lot of jobs, but I don't think it's necessarily negative, nor do I think it's bad for Atlassian. I think it's great for Atlassian. So, we want to help be a part of that agentic world, but the experiences of how we do that is a long way from solved and settled in enterprise software.

Yes, I'm well on record. Headless is brainless. By definition, look, I think again we're confused.

Absolutely. Because when people talk about headless in that way, if you're giving up on your user interface, if you're giving up on evolving your experience for a customer to solve the problems that you set out to solve for them, if you are not a significant part of that experience, yes, I think that's a brainless bet. That does not mean we're not big on MCP. We're not big on CLIs. I think we have one of the top five, top three MCP servers in the world. Like, we're hugely big on MCP and its ability to allow agents and harnesses to contribute data to our platform and to the teamwork graph and to read data from the teamwork graph. And it shows off, to be honest, MCP and harnesses show off the results of our graph, right? How great quality our context graph is to solve all manner of problems for our customers. And that's great. But if our interfaces and our design is not evolving to solve some problems or some significant part of the problem set better for the customer, then yes, I think you're going to be in trouble.

Absolutely. I think it becomes more important. I think design and experience is one of the keys to success in the AI era. Like, we are massively invested in design and experience at the moment in a way that we never have been before, and we've always been quite a large investor in design experience. We have continued to upgrade our experience in a way that I would say few enterprise software vendors have done over the last 20 years. It's really easy to make an app, I always think, with the current, you know, say better than 90% of apps out there today, that's easy. But in five years' time, that app will be better than 20%. You know, experience degrades as we have higher and higher standards. To upgrade app experiences is really hard. It takes a lot of work. It takes a lot of R&D. It takes a lot of difficult craft work to continue to do that. We're in an era where we're all going to need to uplevel our experiences to handle the AI era. But I do think design experience is even more important in the AI era. I think it will be the differentiator. If you argue that technology creation is increasingly cheap, technology creation is abundant and verdant, the differentiators will, to me, clearly one of them will be design and experience. It will be what makes you stand out as an application or as a product or a service.

I think it's a too simplistic view of the world. I think we will certainly have some horizontal agents that handle a breadth of tasks. I think you will also have vertical agents that go very deep in a singular task. I don't think agents will replace all the interfaces. I think all three of these can be true at the same time. So, as an example, we have lots of agents that run in a lot of platforms from, you know, obviously Claude and Codex and Cursor and places like this, and Teams and Slack, that access the Atlassian platform and contribute data back to the Atlassian platform. That does not mean that's the best interface to do all the jobs on the Atlassian platform. It can be very useful for a lot of different things. And vice versa. We're announcing a lot of features with our partner at Figma, for example, and another set of features with our partner Databricks. In both cases, allowing you to get to Figma functionality and Databricks functionality from within the Atlassian platform is really powerful. If I'm reading a Confluence page and I want to just select a piece of text and say, 'Can you check this number for me?' and it goes off to Databricks or to Snowflake through the teamwork graph's data context and gets you back the answer or gives you a small chart, that's great. That's not saying Confluence is a BI tool. Like, it's not going to be the power of what you need. But sometimes when you want a quick answer retrieved from somewhere else, it is a very different experience than if I want to live in that world doing that thing. Same with Figma as an example, right? You can interact with the Figma agent inside of the Atlassian platform in lots of places. You can interact with the Atlassian platform from inside the Figma interface in a lot of places. It's the definition of A2A. I think it's one of the best examples of A2A across multi-craft teams, but both those interfaces have strengths in their area of doing things, right? I don't think there'll be one Uber agent that sort of just does everything magically. I'm a big bear on the magic button thesis of AI, that there'll be one button we press and it does it all for us. I don't think that's how it actually is going to work.

Well, one of the top handful would certainly be continuing to invest and have the best enterprise context graph in the world, continuing to stay ahead. I do think what we talked about before, continuing to evolve the design and experience in the agentic era is going to be a really interesting challenge. Like, I love spending time with our design team at the moment because for the first time in a long time we're solving foundational design challenges again, right? I would say the mobile era is the last time where we had to think about completely new design paradigms. A lot of the features we're announcing here in Amsterdam are really foundational design breakthroughs. How should agents appear in a broad collaboration platform? How should I interact with them? How do I deal with the problems of humans interacting with agents? Agents operating on behalf of humans. Knowing what is AI and what is not AI touched content, and is that a good thing or a bad thing? It's like one thing for AI to 100% generate a piece of text. There's another for it to help you write something, right? And like, how do we explain all this to somebody and where does it matter, right? Provenance and accountability are incredibly important. Speed and understanding. Do I want to see everyone's agent traces? No. But do I need to understand what they are and learn from them? Yes. Lots of foundational design problems that I think will improve customers' results. But that's... it's a really... it's an interesting time for that.

I mean, text boxes suck, firstly. So giving you a little text box and saying, here is a one-line piece of text that you can type characters into and it has unlimited power and it's just a blank white line and you're like, well, that's a terrible user interface to start with, right? And we're trying, we're seeing lots of attempts at changing this problem. It's an incredibly powerful... Agents and chat in general and LLMs are so powerful and most users just stare at them and kind of use them like Google search, right? Like that's the reality of what they do. But how do you explain the power to them? How do you give them the right power at the right times? There's a lot of talk about agentic UI. You know, the sort of word ribbon that changes which functions it gives you at which times and how does it take that as those functions evolve. There's a lot of this that's a challenge, firstly. Secondly, agents doing things on behalf of users is an increasingly common pattern, right? Again, I think software teams are really at the forefront of a lot of this, but I think it'll come to legal teams and HR teams and marketing teams. It's all going to be the same, that, you know, if I'm using Codex and it does something for me, for example, if I'm just writing code and Codex creates a pull request, did Harry do that or did Codex do that or was that Codex on behalf of Harry? How much of that was Harry and how much of it was Codex? Because if you've been heavily involved, it changes the nature of the content and we need to understand where that comes from because people are going to want to see, I just want the human comments here, I want just the agentic comments, and there are valid reasons to see both. This is a really interesting design challenge, right? The provenance of where content has come from is a non-simple question, right? We see this with Anthropic's watermarking features. It implies that watermarked text in AI is bad. That's not true, right? The question is, did I contribute one or two ideas? Did I say, 'Hey, write me an essay about agentic design,' and it wrote me 3,000 words, or did I say, 'Here's my essay on agentic design. Help me make it better,' and it just helps me refine it? Completely different. Same set of text, same watermark. How do we help you solve this problem, right? And then you have a lot of business processes. Every company is just a system of processes put together, a big organic system that's constantly trying to improve its processes. How are we thinking about improving the process versus working on what is flowing through the process and how do we evolve how we run that? We see a lot of this in the software factory thinking that is very prevalent at the moment. And it's not a great analogy, I think, always, because a factory is a very mechanistic, defined-edge thing and software is fundamentally very creative, right? But the idea of it is I'm working on the factory as much as I am watching what's going through the factory, but I'm not stirring or chopping or doing the things anymore or stamping out steel, right? Machines do that, but I still need to operate the factory, I still need to initiate, I still need to handle exceptions, I still need to steer and co-create and co-guide. Often the interfaces for when I jump in and when I don't is a real challenge, right? As there are more things going on in a company, how do I have confidence that when things go wrong I will be notified and when they're going right I'm not notified, but if I need to get into the detail I can, but I don't have to see all the detail because there's just a sea of things going on. So there's so much of this that's... it's just really... it's really exciting, right? A field of change in design and experience right now.

When we started selling software, you wouldn't believe it. We had a fax machine and you would download a PDF off our website. It was a PDF. I think we had a Word version as well. You could download the Word document or the PDF. You would write in your credit card number and you would fax it to us and we would come in in the morning with these rolls of paper on the ground and we would punch it into a physical machine. That is how we originally put payments through. Scott, get off your honeymoon, dude. We got another one. We never shipped a disc ever. So we were like in the internet era, but still payments back then were faxed numbers. That's how companies wanted to pay you. And you would bang it into a physical machine, like type the number in.

I think in five years' time, the whole, maybe the watermarked AI, I think there's a good chance that the whole watermarked AI, the reverse of what you said, is probably true. We don't know what is un-AI-touched content and we also don't think that is good or bad. We don't attribute quality or subjectivity to AI-affected content, AI-improved content, AI in any form. Just like today when you go to the movies, we don't say, 'I want to see films that aren't color-graded.' Like, it's not a thing. You just go and watch a film. You want a great film. There's CGI in all sorts of films. You don't expect, you know, they film in Sydney at nighttime and they make it daytime with CGI and stuff. And does it matter to me? I don't think so, really, right? I care about the quality of the story. I think whether text is, you know, how many people are like, 'I don't want to read your content if Word helped me with the spell-checking and grammar-checking'? There's some version of that that's similar. I guess there's some version where it's entirely AI-generated slop. And so it's like, how will we think about this spectrum? I suspect we all just get used to dealing with the fact that it may be entirely synthetic content and hence artfully crafted things will stand out more. I still think we'll probably have human judgment that makes things stand out.

Jason's increasingly trying to run one of the smallest companies in the world, I think. So, for good or bad, I always find his takes amusing. I think that is something that will probably evolve. We want to be very customer-led in our decision-making and there is certainly some customer pushback in some places to quote-unquote Chinese models. There is not customer pushback to open models. We use a lot of Neotron in our platform. For example, we have a deep partnership with Nvidia and that is fine, right? We use some Mistral, we use a bunch of other open models in different parts. Our job, again, is to apply AI on behalf of our customers. It's not to pick a model. It's to use the best model for the best job, blending all the models for the best price for the customer and to get the outcome they want, right? Applying the AI is our job. But if there is customer pushback to something like a Chinese model, quote-unquote, that is something that we have to be very careful of. And so we don't use any Chinese models in production, we're very clear about that. I think what people don't understand is the distillation and the back and forth where knowing it's a Chinese model is kind of really interesting, but a lot of the US models are increasingly using Chinese models to train themselves, but they don't say that. And so you end up with all these weird loops where I'm like, I don't know if that disappears over time. I suspect it does, but I think we'll have to see and watch. I do talk to customers sometimes and say, 'Do you use Linux?' And they say, 'Yes.' And I say, 'Is that European?' And they say, 'Yes.' And I said, 'Well, I mean, I think it's like 40% of contributions or something like this is quite high nowadays come from China because there are some amazing software engineers in China doing all sorts of things, but because it's open source we think, oh, well, someone's like checking all of that.' And actually, weirdly, with AI we can check that even better than we ever have been able to before, and they're happy to use Linux, so...

I think there's a geopolitical issue around who has this sort of strategic control of the... That is a different issue from should we use a model that was trained in China or by Chinese nationals or by... like, some version of these two are very different questions, if you know what I mean, in terms of who's running the world, America versus China. That's a big question that's out of my league and my...

But I'm not commenting on any of the social issues or other issues in that area. What I'm saying is, from a technological perspective, for someone to say, 'I'm not going to use a piece of technology that is open,' right? If I'm saying, 'I'm not going to use this piece of open source code,' I'm like, 'But I have the code.' Well, if you choose not to use it, that's totally fine. That's your choice. But for an enterprise to say, 'I'm not going to use a cheaper alternative that'll give me better results when I know what's in it,' I think is going to be a very hard long-term choice. And I suspect people are going to take those open-weight models, which they are already doing, and fine-tune and train them and call them their own. Most of the people who are training their own model, you know, large company X claims they've built an amazing model for purpose Y and it's got some wacky name. We've seen many, many times, almost every single time, that is, if it's not Neotron, which it's usually not, it is a Chinese model that they've trained or added their own fine-tuning specialization on top of, and then they call it their own model. And I'm like, hang on, that is the same thing, right? You've added to it and you've tweaked it. You haven't trained it from scratch.

The jury is really out on the benefits of building specialized models at the moment. There are definitely cases where specialized models are doing really well. Most of the people building specialized models, I would argue, are not in those cases, though. There are certainly, if you have a lot of proprietary data, if you have very specialized applications, there are definitely approaches to that, right? So if you're doing, you know, certain robotics, for example, and world models, there's a lot of reasons why people are actually doing it that make sense to me. There's a lot of others that I don't think will stand the test of time. They may have, at the time of shipping, a price-performance benefit for a specific application. True. Will that price-performance hold up over time? I'm not sure it will as generalized models get better and better and better for a lot of applications. Firstly. Secondly, a lot of people who say they've trained their own model, I think they're being a little bit fast and loose with words, as far as I'm concerned. We could say we've trained a bunch of models at Atlassian. I say that's... We have a bunch of very specialized open-weight models that we fine-tune for particular purposes that give great price-performance for our customers. That I don't claim to be our own model. I would say like, that's where I think the claim is often vastly overweighted. Most of these people are taking a Chinese model, generally applying some domain-specific intelligence. You see they train them for like a million bucks and I'm like, there's no way you trained this for a million bucks. You took someone's billion-dollar model and tailored it, if you want to think about it generically, for a particular application. That's totally fine, but that's not really training your own model, I would say. Right? If you're not starting with raw data. If that's their definition, then we do that a bunch in a bunch of different places, but it's not my definition of building a model from scratch.

If they could continue to show exponential gains over and above using the latest foundational model on a sort of price-per-watt basis. So we are seeing many more applications of... and it's very hard to answer these questions generically because it depends on the target application actually, and the time to answer or solve a problem or a question, the cost, and the period of time you're in. If you look at an agentic harness, it is a way of applying a set of prompts and configurations, if you want to think about them that way, to a certain type of model and reapplying it many, many times, if you want to think about it as looping, so that you can get much more value out of a given class of model. And yes, if you apply better models, you get better results from a lot of those harnesses. And harness design is really, really important nowadays, but it's not model progress. These specialized models have to prove that they can stay ahead of the generified model for a given price-performance ratio. I don't think I see many of them doing that. Certainly not the vast majority of them doing that, because you would need to have a continued specialized dataset and a problem that doesn't get generified away. Many people who I think had startups, hopefully none of whom you've invested in, that were building their own models two or three years ago, I think now the foundation models have just basically smashed what they do, right? Like, the quality of the generic model has improved far faster than their ability to stay ahead with whatever specialization they were doing. So I think I would need to see the reverse of that.

Amanda, I love all these private companies that have raised a bunch of money and don't have to care about cost because they're not building the discipline to care about cost. Some of them do, but over time, discipline is really important. Anything you do in life, it's about discipline at some level, right?

I guess that depends if you continue to fund them and believe they'll be funded forever. It seems unlikely given the history of technology, but maybe. I think cost is really important, but it's not just because of that. I think if you look at a lot of examples out there, it's very easy to use a very expensive model to get pretty good results pretty fast. I think using a cheaper model or an older model and getting good at applying that to problem sets is a real advantage for most technologies and prices and costs. I think customers that we spoke to, so large enterprises around the world, they're incredibly sensitive to this type of stuff nowadays. They really, really care about the price-performance. They are increasingly doing so. We have a whole raft of products now that will show you your token spend, either across the board in terms of your talent, either in DX, in your engineering organization, right? Understanding where you use which model, where you're getting value, which features you're using it for, where you should change to optimize your spending, because this is a bigger and bigger budget item and they see it's only going to be a bigger budget item and so they have to care about the spending. Now, that I don't think is any different for them buying a generic model or a foundational model off any of the big three or four or five as it is buying a product where those costs are going to be baked into whatever that product they're buying is. So I think that cost discipline is going to arrive continually more and more. And I think it's a strength of ours, for example, in how we are able to use our R&D strength to continue to reduce the price of features that are used heavily. I think that is a core strength and attribute of a great technology company will be its ability to manage technological cost.

Model routing, model blending, applying the right model at the right time, using the right harness that again picks and adjusts itself. It's a real verdant area of exploration right now, right? And we... you have this... the more you can get good at it, the more again you can apply AI to smaller functional areas. If you look at the so-called decision models, Jeb and the 15 clones that have emerged over the last three weeks, which is insane and amazing and scary all at the same time, they are often used for tasks that previously would have been too expensive to use even like a GPT-3 or GPT-4 level...

model for, and that's great, right? Because we're now applying probabilistic thinking to more and more feature sets. I think we're all just going to see better products as a result.

I think for companies of our scale, you have to be very good at doing that yourself. There is a constant challenge with scale. I think any large technology company is probably going to do that internally. And that may involve using OpenRouter plus many other things, plus your internal router and everything else, right? Like these things can be continually layered upwards. I think for smaller businesses or businesses where technology is maybe not their expertise, those are just very practical solutions, right? There are many, many model routers and gateways with lots of claims about what they'll do. I think generically, it's very hard to just route automatically, I think is what they're showing. You need to give enough steer or intent, and I think harnesses will also be a part of this. So this is, you're sort of talking for generic LLM calls. The more you're using harnesses, they're doing their own sort of routing, right? The more you're using SaaS, you know, you come to the Atlassian platform and you use Rovo or any of our AI features, you will get model routing happening for you automatically. And it's our job to blend those costs and give them back to the customer for the best quality answer, for the cheapest quality price at the best speed. This triangle trade-off, we're always looking at on any given feature. That's part of our USP to the customer.

That is all internal routing. Yes.

We have an entire AI gateway team. It's not a small team that is doing that. But we have to do it firstly at massive scale. And secondly, we want to continue to reduce that cost where it makes sense. So I say to product teams all the time, don't overoptimize for cost in the short term. Build amazing magical features. If they're taken up by customers and what we thought was magical, the customer thinks is magical and it scales, now let's start working about the cost. Right? So think a little bit in advance. Can we possibly get this cost down? But then worry about that later. Right? If five people are using a feature, I don't care how much it costs. If five million people use a feature, I'm going to care a lot about cost. And in the middle, I usually have a fair bit of time to keep optimizing.

Model inference. My maybe slightly spicy take is that the amount of energy, startups, capital, time, smarts, brains, people who would otherwise have gone to work for McKinsey or NASA or something else, working on some version of the price per watt performance of inference is crazy. And I suspect a lot of those take years to come through, from new chips to better data center design to ways of distilling and getting things out of foundational models. Like there's lots of the overall progress that I can't believe doesn't bring our price performance unit of an LLM, we don't have a good term for this, down at a precipitous rate just because the volume of capital being applied, the volume of brain and talent being applied is so large.

We will, but it will still mean the cost comes down. So where the cost is near free, yes, it will explode. If something's free, you use an unlimited amount of it, right? And you don't care, you throw it away, whatever. Like if it's not got a waste component to it, which is ostensibly a cost, then yes, that will be the case. I think there's always going to be an economic curve to that, though. Right now, the production of technology is faster than the consumption of technology, right? It is ahead. We've reached a tipping point. We can build features faster than customers can consume them in a realistic way that changes their business. There's got to be a point where that just, it doesn't stack up anymore, right? There's not going to be an infinite usage of all of this stuff, right? Your phone has amazing superpowers. It's a supercomputer in your pocket. Do you burn the CPU 100% all the time? No, you don't. You're not even close. Right.

Look, I think it's just incredibly positive, I suspect, for those who want to use those models to achieve things, whether you're coding, doing architecture, improving business processes, whatever it is that you are doing, creating art. I think we should see that as a very positive thing. We are going to get higher value technology. We are going to get better products and services for every consumer. I think it is going to be a creativity-driven economy. One of the phrases that I've liked the most recently, and I give Didier Elzinga, the founder of Culture Amp, the credit because it's his phrase, I believe, or at least he's the first person that told me, we're moving from a world of information asymmetry to imagination asymmetry, which is such a great way of thinking about it, right? It's no longer I know something that someone else doesn't or I have a particular business knowledge. Those knowledges are becoming less about the asymmetric value than can I think about how to apply AI and these technologies to problems in unique ways. That's the creativity that we're all kind of having to see in business now, which is if you're a first-principles thinker in our business, it's awesome, right? It's the most dynamic and exciting period that we've ever lived through in technology and maybe ever will. I get why it's scary. I also think it's imagination. You're applying new things in fundamentally new ways, which is...