How AI Agents Are Transforming Workflows Through Intelligent Automation with Jamil Valliani

Hey everybody. We're back at Atlassian Europe and we're here with Jamil, who's the head of product for, uh, AI for all things at Atlassian. And we're gonna have a little chat about, well, where is AI headed?

Jamil, welcome the show. Yeah, thanks for having me here. We seem to have gone from co-pilots to AI agents in a blink of an eye, and now we have all these helpers that are gonna do all kinds of work for us, but where is AI headed from here?

How smart can smart gap. Yeah. I, I think that a, a lot of the future now is not gonna be necessarily about, you know, who has the, the best model.

That certainly is important piece of the puzzle, uh, but really on how, uh, people are bringing that, uh, capability into their day-to-day work, into their workflows and making it easy to access, easy to make, and, and part of their teams. And that's where we're really focused right now. Uh, we found that a lot of the you success or failure of our customers in AI doesn't have to do with, oh, is the model doing the right thing?

Or does the software exactly the perfect for what that task is, but it's in how effectively it's integrating into their workflows. Right? So one of the things about AI is it's probabilistic.

And so it's giving you, you essentially its best guess of what you need. Yeah. Um, we have a lot of processes that are deterministic sometimes where it has to be done the same way every time.

And, you know, that stuff's kinda wrote, but do we need to be careful about how we're thinking about applying AI based on what the actual mission is? And we gotta have to back it up from there. I mean, 'cause I think a lot of people don't think about the nature of the workflow all these day years.

It's one thing to create an email, it's another thing to process a contract. Absolutely. I, I think that there are a lot of deterministic processes and they've already, there are already tools out there, uh, like automation platforms and whatnot that are really good at trying to simplify that.

But what we find is that, uh, folks who are trying to plug in agents into automation flows, for example, what they are trying to do is actually make that automation more robust. Because at some point there is some judgment required, and oftentimes that judgment is, you know, fairly straightforward. Uh, it's something that requires a bit of context, uh, a bit of understanding of, of, uh, you know, the content, the rules, all those sorts of things.

And then it can apply some sort of judgment. And we find that in that case, uh, those are like prime examples where an, um, an agent can actually start making a workflow better, uh, that was previously deterministic and only able to handle a certain amount of capability, uh, that I think we'll see a lot more of in the future where those, like, you know, couple layers of initial judgment that that triage step, for example, will start getting assisted by agents with human review, uh, but will still simplify a lot of the work that has to get done. Um, if you think about, for example, analyzing a, a Jira backflow, uh, you might have like five or six themes that you thought of, and it's a very simple judgment to say, well, which theme should it go into?

But nobody really wants to like go and spend their whole day doing that. But if you have an agent saying, Hey, like that, that is a, a sort of istic problem, but does require a little bit of judgment prime a candidate for something that we would actually accelerate with ai. Yeah.

Won't we also use AI agents to kind of review the work of other AI agents at some point? I mean, we talk about human in the middle, but maybe I don't always wanna do that myself either. So will there not be times when I'm use an AI agent to kinda look at what some other AI agent did and make sure it's not hallucinating and put some guardrails in place?

Absolutely. I, I, I think that, uh, these sorts of systems will become pretty commonplace. Uh, and I, I'm, I'm excited about the innovation that's happening there already.

Um, a lot of the initial wave of, uh, review for AI happened as agents were starting to get plugged in to review, uh, labeling, for example, like one system would label, the other system would judge, uh, even in engineering right now, when we evaluate VO Chat, for example, we'll actually go and say, Hey, VO Chat gave this output, let's go and actually have a separate, uh, agent that runs, that helps us grade that output later on so we can evaluate if you think it actually answered the question or not. Mm-hmm. So what is that line between, I mean, one of the things that's become apparent is rovos everywhere in the portfolio.

Yeah. So what is the line between Robo a product and robo a feature of something else? Yeah.

So we've, uh, we've to said that RO o is part of our platform, uh, and we've considered that the platform offering of robo to have three core applications, uh, search, chat, and studio. And we believe that those core applications are gonna be the basis for almost any teamwork in the future, regardless of what discipline you're own, everyone needs to search. Everyone will want to go and actually query their knowledge sources and take actions.

Everyone would want to go and build a hub, automate their flows. Uh, so we believe that's like woven into the platform that's the right thing. And then for every collection that we have, every app that we build, uh, our intention is to build a set of, uh, you know, tailored agents for that particular, uh, collection, uh, as well as specific applications and enhancements that plug into the robo platform.

Uh, and that again, will be like very tailored to the, you know, cases and, and needs of those customers who use those tools. And where is that line gonna be between the agents that you provide and the ones that I might go build using Studio? What are, what are some of the use cases look like?

Totally. Yeah. I think that there, there'll be some agents that we provide that are ultra sophisticated.

So for example, we announced Rob Oev, uh, and Robo Dev is a, a code generation agent that is actually able to harness all of the context, uh, of, you know, code and the software building that we have from you, because you trust us with your Bitbucket, with your, uh, JIRA and Confluence. We can do a really good job with that generating code for you. Uh, and that's a very sophisticated agent.

So we'll provide that. Um, the other agents we provide that are really meant to be examples of starters, right? Uh, like we won't presume to know what the best way is to go in, uh, you know, triage every, uh, or prioritize every bubble list for someone, but we can provide you with a charter, right?

And then we wanna encourage customers to go customize that agent for their needs. And we hope that that leads and inspires folks to go and build even more agents that are tailored to their business processes based on the examples we provide. And then we continue providing these really sophisticated ones that, uh, really require, you know, more than, uh, typical producting.

What is the future of the software user experience gonna be like? Because historically we've had all these tools and different GUIs, is AI gonna harmonize all of that? And maybe I won't even know when I'm in and out of a given product.

Yeah, I I think that we're in the very early like, you know, dos command line era, uh, of how we interact with ai, right? Uh, and, and that's totally understandable. I I think that, you know, typically people start with these, you know, simplistic chat interfaces.

That's a very natural, easy thing to understand. But as AI is solving more and more problems, uh, and people get also more and more comfortable with a predictive model helping govern their application experience, I think we will be innovating as an industry new, uh, experience paradigms that fit naturally into that. Uh, I think we're still in early days on that, but, uh, I think with what we're doing, for example, with, um, the browser company, uh, and the opportunity, there is a good example of a, a place where we're actually making an experience bet saying, Hey, if we had to go and think about an AI forward way, um, of actually using your, not your SaaS apps every day, what would it look like?

Um, that's one example of a bet we're making, and I think the industry is gonna know push in that direction on there. One of the things you announced at the show was support for the model context protocol. I think it's coming early next year.

Yeah. Um, when, how will these AI agents kind of interoperate with each other? How will we orchestrate them?

How do you see that whole thing evolving? Yeah. So we fundamentally believe in, um, an open platform.

Uh, and that's part of Atlassian's DNA. We think that we work better together with other partners, with third party vendors, uh, and that that's what our customers expect. So, uh, we actually already have our MCP server out there.

People have share user server, and then, uh, we're gonna support MCP services part of studio, uh, very soon as part of the studio release. And that's just showing, um, our commitment to that, uh, to that principle. Uh, I believe that there's no one company ever that will have, um, total understanding of every business process and every need for every company.

And that by nature necessitates, um, this, and, and sets the expectation that we provide as an industry open platform so that people, vendors who have that specialization, who have that skill can provide those tools. And that we interoperate together. We announced also partnership with Google, for example, where we'll work with them on agent to agent.

Uh, and that's another example of where we're saying, Hey, we, we will, we know that we're not gonna go and solve every person's problem, but Google will solve some problems. Other vendors will solve some problems, and we'll have to learn how to operate together. And that's where our mind is at on this front.

Mm-hmm. So in that environment, we may have, let's say I have a couple AI agents, and you have a couple of AI agents. Will they negotiate with each other?

How will that discussion kind of evolve? Yeah, it's, uh, it's a good question. I think a lot of innovation is still happening on this front.

Uh, what I expect will happen is that, you know, every, uh, agent out there will have to have a responsibility that's pretty significant to manage the trust and security, um, of the content that they have responsibility to govern. Uh, so for example, when somebody, uh, connects to the VO agent right from another platform, uh, you know, we're governing access to the data on the Atlassian platform that, that, that's being requested. And it's up up to us to, you know, understand and work with the other vendor to, you know, have the right permission structure in place so that we don't accidentally leak data.

That, that the right data goes back and forth and decide break customer plus expectations. That's a very important step that it has to be a part of that, you know, protocol and negotiation, um, that I think is easily underestimated, but very critical. Uh, I think the other thing that's gonna happen a lot of is this idea of orchestration, right?

Who's actually owning this decision of when to actually call out to the other agent? What is the base on which that decision is made? Uh, that's gonna be a very powerful, um, capability and an important one.

We think that we have a, a very good position there because people trust, uh, Atlassian and, uh, put their knowledge of their goals, their projects, their work items, et cetera, all with us. That gives us a lot of context to which to know, uh, how to make that kind of decision, which will, I think, be a, a very important pillar of these protocols and featuring, is there gonna be some sort of hierarchy of AI agents? 'cause I don't think they're all gonna be created equal.

And, you know, might this evolve into something that feels like, you know, upstairs, downstairs there's a head butler and then there's a bunch of agents in the basement doing interesting things that are relevant, but nobody ever sees them. Yeah. I, I think, um, I, I'm not sure if it'll play out quite that way.

I think it's a, it's a bit, it's easy to pontificate. It's very hard to forecast right now. Um, I think our belief is that, you know, every, um, you know, every major app vendor out there will have a set of agents who are a singular agent that, uh, you know, serves the task that they have.

Well, uh, we certainly would think we have a lot of things that we can offer with robo and, and unique capabilities that we have, um, as part of the Atlassian platform. Um, but we also fundamentally believe that there'll be other agents we have to work with. Um, I don't expect that there'll be one agent to rule them all.

Uh, I don't think that's how it works. Like, you know, we've had even these generation one assistance around for a long time now with whether it's Siri or Alexa or Cortana. Like, none of these have become like the single dominating agent, and customers don't seem to want that.

Mm-hmm. Uh, and I think that's reasonable, right? I, I, I don't think that, uh, you know, people are really comfortable yet with just having one personality that discovers everything.

Am I going to develop a relationship with my AI agents? I mean, or will they just be kinda like, you know, things that pop up every now and again, like, you know, bad example probably, but clippy. But, um, or is it gonna be something that, you know, there's gonna be some entity that I recognize.

Maybe there'll be a few of them, but there'll be something that I will put a name on. Yeah. Yeah.

I think that at, at a minimum, uh, people will expect the agents to understand their personality and their preferences. Uh, one of the things that we announced this week was this idea of personal memory to compliment the organizational memory. And that's because a lot of the things that people expect on other agents are nuance.

Uh, but but important for them to meet their needs to, for example, remembering that I prefer long form content while somebody else prefers bullets and emojis, um, is a personal preference. Uh, the AI has to learn and understand and preserve about you. Uh, and I think the more it's able to pick up and understand those personal preferences and those, uh, those sorts of details, the better it'll be at serving you.

Um, that doesn't mean full on personality. I think that, you know, we'll have to see how things evolve and if people prefer that. But that idea of, you know, having this, you know, deeper knowledge of, of you at that level, we'll just make decisions more powerful and rely on it for people.

And we realize this is something of a more subtle shift, but yeah, when we had copilots, people were studying up on prompt engineering. Yeah. And now you hear the phrase context engineering.

So is this whole space gonna evolve and change? 'cause maybe I'm not gonna be the master of prompts as much as I'm gonna be figuring out what context to give the AI agent, but I gotta give that context in the right order. Uh, I think both will matter.

Its tongue. Uh, and I think what folks have been learning is that, um, a lot of times you can do all the prompt engineering in the world, but if you don't feed the right data, uh, you're not gonna wind up with the output output that you want. Um, and so I think, uh, I don't think we'll see like one be more important than the other.

I think that as people get more and more experienced building these agents and these sort of AI capabilities, there'll be best practices that build up around how to set them up for success. Uh, that involves how to write the right prompts, how to feed the right context in. It also involves things like where to weave it into your workflow.

Um, how does it actually interact with the team? Like is it a, a distinct identity? Uh, does it get permissions?

You have to give it skills. These are all, uh, things that are being lured now in real time and that we're rapidly incorporating into our studio, um, to make sure that customers have the ability to, to dial those things in and, and set their preferences up. Uh, and then a lot of testability.

And, and trust also has to have factored in. Like, knowing that an agent has a good track record actually is really important. Uh, and I think we'll see more systems like that put into place as well.

Tell people who these things and start trusting them and do real life. One of the things that I think is not so much a secret out there, but a lot of our processes are, shall we say, not very neat. Yeah.

Will AI and the agents kind of force us to kind of clean that up a little bit because they're gonna be looking for, you know, more structured things and the more structured give it, the more context it has and it becomes a virtuous cycle. But are we gonna have to go revisit a lot of our processes? You know, I think that we will have to revisit processes, um, but I don't think necessarily to make them more structured.

Um, I think that that's where, um, I would expect the agent is meeting you, right? And say, Hey, like, let me go and take this challenge anything off your plate and run it better for you. Um, but I think people, excuse me, will have to learn about where is the best place, what are the types of best problems in a, to trust an agent with?

Um, that's what I think take a bit of muscle building, uh, and also a bit of work from technical teams as well to sort of figure out how, how good we can make these agents to adapt to different types of scenarios. And then people will learn, I think, hey, like, here are the types of scenarios an agent can learn well, uh, work well. Um, so I think we'll, we'll see a, a lot of like learning from both, you know, the, the workers and the agents on, on that front.

Um, but I expect that the problem people will wanna solve is actually what you mentioned is like, Hey, I have like a really challenging process. Can you help me go and, you know, make it run more reliably, more smoothly, think more predictability, um, and I think actually lead to better, better processes. Will the AI agents also be able to surface the dependencies that exist between processes?

'cause I think we try to keep all this stuff in our head, but um, we sometimes forget that, you know, five projects are dependent upon this other project in on time and that project's late, but nobody knows. Oh, absolutely. I think that's like one of the best initial use cases of these agents is often, um, you know, you have so many different knowledge sources and pieces of data out there, and it's really helpful when you actually are able to trust an agent or even chat to go and, you know, pull all those different things, you know, suss out the right relevance of information, provide the reference links for you so you can follow up.

Uh, but, but see where which ones are worth spending your time on. Um, I think you can do that today, and I think that's actually a super value to, you've been at this a while and a lot of other folks are kind of newbies, but yeah. Is there something, you know, now that you kind of wish you knew a couple of years ago?

Oh, wow. Uh, so many things. Uh, I'd say, uh, a couple of the, the things that I'd share with, with listeners, um, I think most important is to start out with actually, uh, something small, right?

Uh, yeah, I think there's a lot of temptation when you, like say, I'm making an AI investment. Let's go solve these giant problems and certainly have the ambition. Uh, but we find that the best success stories are often when, uh, folks are able to start off with, um, just one or two pain points, very discreet pain points in their system, in their processes, in their teams, and we're able to go and say, Hey, how do I go and, uh, make that even 10, 20% better?

Uh, we find that if you're able to go and focus on a, a narrow problem, that's where AI can do the best job, uh, right off the bat. And then you can start expanding from there because you'll learn the AI will learn as well. Your prompts will get better, your conscious will it better.

Um, you'll be able to be run more evals. All those things will get better and better. Uh, we actually reach this ability called of scenarios, and that's sort of in the same thought process as saying, Hey, try with one scenario, then you can add second, third, fourth, fifth scenarios.

But we find that teams that go in that, that manner, uh, tend to have a lot more success. The only other thing that I'd, I'd share that, that we've learned a lot of is that there's a, you know, a culture shift as well that's important. Uh, and it's both top down and bottoms up.

It's very important for the leaders of every company, uh, to really lean in and say, Hey, I'm gonna go try these things myself. I'm gonna share my successes and my failures with my team, uh, and show that I'm being vulnerable and, and trying these things out that sets the tone for everyone else to, you know, be comfortable. And then from a bottoms up perspective, there's always a few teams at every company that, uh, are ready to take risks that are, are most risk oriented, want to go and take a chance, take a plunge, have a big problem to go solve, and it's important that they get supported.

Um, and I think it's easy in a big company, um, to have a team that's like, that feel like they just don't have the room to navigate and score. But I think it's important to actually give those teams room because they tend to find those solutions and then set the stage for everyone else to, to fund. So, all right, folks.

You heard in here, even in the age of ai, you still need to learn to walk before you run. Yeah. Hey buddy, thanks for coming by.

Thank you very much for having me. And I enjoy the time here in Barcelona. Thank you.

Yeah. And we will be back in a minute.