GitLab’s Manav Khurana: AI Agents, Orbit, and the Future of Coding

[00:00:00]Hi, I'm James Magcguire and on today's Tech Voices, we're talking about software development and artificial intelligence and we're focusing on topics like the life cycle context graph, building AI agents and the future of software development. We're going to talk about where it's all going. To discuss that, I'm joined by a major industry expert. With me is Manav Kirana, the chief product and marketing officer at GitLab. Manov, really good to be talking with you today.

[00:00:26]>> Hey, James, so nice to be here.

[00:00:28]>> Great. So I think there's something really interesting about AI and development in that uh developers report being 20% to 50% more productive with AI tools. No surprise there. But companies aren't really seeing the same gains at the business level. What's going on there? What's the discrepancy?

[00:00:47]>> Yeah, I mean I confirm both those points. You know first it's clear that we are in the agentic engineering era >> definitely >> where uh we're finding you know when we survey our customers survey developers we're finding uh 90% plus uh of developers are using two or more tools >> right >> reporting much higher uh productivity and speed uh in their daily jobs. Uh in fact we've seen uh our customer code bases grow as much as five times over year. Uh >> oh, did you say over the last 12 months or >> That's right.

[00:01:20]>> Okay.

[00:01:20]>> That's right.

[00:01:21]>> Wow.

[00:01:22]>> Uh and um you know, but the thing is like you were saying in the second part of your uh your note there is um but that speed is also bringing chaos. M >> we see the chaos in terms of uh reliability incidents in the industry or security incidents in the industry or um agents hallucinating and showing up with artificial confidence. Um and that you know leaves it on all of us to then figure out where did the agents go wrong and that's what's causing uh enterprises and businesses from not realizing the value even though the individual >> is the value >> right so that the artificial confidence andor the the chaos there is causing a problem I guess the real question is what does it take to close that gap >> yeah you know so uh we've been focused on that specific problem statement um you know even Before the agentic era, our customers have been using GitLab as their control layer >> for every software development task, right? Whether it's a human writing code or doing a software development task or an agent writing code, all of that tends typically go through GitLab where they make sure the code is of the right quality. It's at the right security, it meets all their internal policies and compliance goals. And we have extended that same platform for agents as well.

[00:02:46]So that enterprises can get agentic speed >> with enterprise control.

[00:02:52]>> Uhhuh. Interesting. Well, I think we we know also that that AI agents are generating code at an enormous volume and speed just really like never before.

[00:03:01]I think in many cases traditional development infrastructure was never built for that. Um how does GitLab approach that problem and and what role does orbit play in keeping that scale?

[00:03:11]>> Yeah. Uh I actually think so before we dive into uh orbit I think maybe it's worthwhile explaining our our platform and what we how do we ma provide that control.

[00:03:20]>> Yeah let's back up and do that definitely.

[00:03:21]>> Yeah so you know the way I think about this is GitLab as a platform is the motor system immune system and nervous system for the human brain and the agentic brain doing software development tasks. Ah okay.

[00:03:38]>> Motor system as in the arms and legs, the execution layer, >> right?

[00:03:42]>> That helps you write code, ship code, test code and the like.

[00:03:47]>> Mhm.

[00:03:48]>> Immune system as in the governance layer that tracks every action whether it's done by humans or agents and make sure the wrong action is not being done. and the nervous system which provides humans and agents the right information so they can make a better decision faster.

[00:04:07]>> Okay.

[00:04:07]>> Right. Those are kind of the three core components and then there's a layer of agentic automation on top uh that GitLab provides today. Uh back to your question then uh yes it is true that the infrastructure is under pressure because of the amount of code being written and the number of concurrent sessions that are happening of writing code and hitting the underlying infrastructure >> right >> and uh what we're announcing uh at transcend is uh a new next generation git backend that uh operates at about a 100x faster scale.

[00:04:48]>> Wow.

[00:04:48]>> When there are uh concurrent sessions hitting the back end.

[00:04:53]>> Okay.

[00:04:54]>> It used to be, you know, when all of us as humans were writing code, right, we would clone code into our local um development environment or we would make changes to our repositories every so often.

[00:05:10]>> Mhm. Uh but now each one of us in the development community has hundreds of agents that are working concurrently.

[00:05:17]>> Sure.

[00:05:18]>> Right. And that extra load is what's push putting pressure on the git back end.

[00:05:23]And with the new next generation git, we have ar rearchitected that backend to handle that load at much much higher scale. And that's why in fact just before our call James I was just checking the uh the system in a sidebyside harness and I found that uh doing 100 concurrent sessions reading information back uh into a local environment was 176 times faster.

[00:05:51]>> Okay.

[00:05:52]>> 3 seconds instead of 10 minutes uh plus to to do that uh same transaction.

[00:05:58]>> So it's really a world apart from the way it was. really a world apart for sure.

[00:06:04]>> I was going to say what what what does that mean for the human developer? How does that change his or her life?

[00:06:08]>> Yeah. So, uh you know, as you know, over the last few months, all of us have been dealing with git uh reliability issues.

[00:06:18]>> Mhm.

[00:06:18]>> Right. Um >> it's been in the news to be sure.

[00:06:21]>> It's it's been in the news. And when the Git back end is not working because it's under load, it's not working for agents, but it's also not working for all of us.

[00:06:30]Mhm.

[00:06:30]>> when we are trying to make changes, right? And that slows us down, that slows my company down and the like. Uh with this higher scale git, not only is it working for agents, but we get consistent reliability and performance that we all expect from a modern system.

[00:06:45]>> Uhhuh.

[00:06:47]Well, what about a use case for orbit?

[00:06:49]What's what's an example of how it might be used?

[00:06:51]>> Yeah. Yeah. So, orbit uh think about it as the nervous system component >> where um so today our customers get you know a unified platform for dev sec ops.

[00:07:04]>> Mhm.

[00:07:04]>> Where uh there's a common underlying data platform data store. So if a developer is searching for hey I'm trying to fix a pipeline what is the related issue that this pipeline is for or what was its prior result? All of that's available, but you still have to kind of stitch it together.

[00:07:25]For our larger enterprise customers, this information is spread across hundreds of thousands of projects.

[00:07:32]>> Mhm.

[00:07:32]>> And it's challenging to stitch it all together as a human and it's certainly challenging for agents to do the same thing because they typically run out of their context window and then start hallucinating or operating on insufficient information.

[00:07:46]Orbit is a digital twin of a GitLab instance that has all the relevant information stitched together.

[00:07:56]So instead of doing 200,000 lookups, you do one lookup and all the relationships are available right away.

[00:08:03]>> H uh and and that's the thing that we're introducing on June 10th in public beta for all of our customers. We've been using it internally as customer zero now for a couple months and having been having a lot of fun with it.

[00:08:18]>> Interesting. Well, when I think about a a digital twin, I'm thinking about something that is typically constructed by, you know, a software architect. H how does Orbit build that digital twin?

[00:08:28]>> Yeah. So, the technology we have built with Orbit, it indexes all the code across all the entire instance our customers have.

[00:08:38]>> Mh. uh indexes uh what that code is doing, the functions it's calling, indexes all related metadata like who wrote that code, what was that code related, what issue was it that code written for, what was the CI pipeline related to it, what were the security test related to it and all that information is stitched together automatically and we make that digital twin available in the same instance that the customer has. So the data is protected and private and only available for them so that their team and their agents can access this and get uh much better insights uh right away.

[00:09:15]>> Uhhuh. What is is it accurate to say that orbit reflects a major shift in enterprise AI from better models to better context or you don't think of it that way?

[00:09:25]>> That's exactly it. That's exactly it.

[00:09:27]You know uh there's an old adage in in the AI world which is AI agents work as well as the context that they're they're given.

[00:09:34]>> Yes.

[00:09:35]>> Really about that context component that um uh that we are fixing. And you know I'll give you a couple of examples here >> please. Uh so uh one of my teammates uh they were using plot code to uh resolve a bug >> and they did it side by side between um without orbit uh putting that information from GitLab and then with Orbit uh putting the same information to fix the bug um with in GitLab, >> right? And uh what they found is that the bot code, the same exact agent worked uh about 40 or 50% faster at that task.

[00:10:21]>> H >> required about 30% fewer tokens uh and was a little bit more accurate when it was when that developer was looking at the actual work being performed uh in that scenario. Right.

[00:10:35]>> Okay.

[00:10:36]>> Right. So it's faster, more accurate and cheaper. And then as you know, we also build uh we also offer Duo agent platform. Uh Duo agents are built inside GitLab.

[00:10:48]>> And uh one of my other colleagues was doing a um uh an analysis of a codebase.

[00:10:56]Uh he was trying to see uh what the log 4j vulnerability status is across the entire GitLab instance. Mhm.

[00:11:04]>> So it's a deep research type uh type use case, >> right?

[00:11:08]>> And and doing that use case without orbit just failed.

[00:11:13]>> Ah >> because one they had to open up thousands of repositories one at a time and they were not able to get the right information.

[00:11:21]>> Right?

[00:11:21]>> Doing that with orbit just took a few minutes and the information was not only more accurate, it was actually a zero to one difference, right? because now it's actually truly possible to to get answers that were not previously possible.

[00:11:34]>> Right. So really the the developers will be will not be waiting for the work to be done. They'll be moving far faster.

[00:11:41]>> Correct.

[00:11:41]>> Yeah.

[00:11:42]>> Correct.

[00:11:43]>> Well, we we know that AI agents are making decisions across the software life cycle with limited human visibility into what they're actually doing. It's it's it's pretty remarkable really.

[00:11:52]We're not sure what the bots are doing sometimes. So, how do how does AI governance change that? I think we talked about this a little bit, but I think it's worth drilling down into. And what does it mean for organizations that need to operate AI within strict regulatory boundaries?

[00:12:05]>> Yep. Yep. I think this is where that immune system part of the GitLab platform comes in. This is about tracking every action and enforcing who can do what actions and which agent can do what actions. Mhm.

[00:12:19]>> Uh so even before um uh today before the June 10th transcend event, we've uh provided our customers the ability to define which agent can do what and also have a unique identity for each agent that follows the the owner's identity.

[00:12:37]So you can uh track it uh and one can define or customers can define which project which codebase agents have access to or not and what kind of access they have uh at any given point.

[00:12:51]>> Uh what's new is uh we have built a a more s a much better way of understanding what agents did with our audit trail capability. Okay, >> we were always tracking what agents were doing. Now we can actually see if what agents did met my policy or not.

[00:13:14]Uh we can see what agents did is violating some compliance rule uh or opening a risk vector that I previously did not appreciate that was out of my consciousness. GitLab surfaces that information directly uh to our to our customers. So that's the first key thing. The second is much more fine grain uh controls on what kind of action, what kind of tool can an agent do by itself or where there must be a human in the loop.

[00:13:51]>> Uh and you know this is a long list of different kinds of actions agents can do that one can define now at a global level, at a project level, at a personal level.

[00:14:01]Well, I'm hearing the the fine grain.

[00:14:03]That's interesting idea and that I think managers are pretty confused or maybe even concerned about AI agent governance. They're thinking, well, a AI can hallucinate, so can't my agents hallucinate? Does that fine grain control address that?

[00:14:18]>> Yes, because we sit between the agent and our customers assets.

[00:14:24]>> Okay.

[00:14:24]>> Right. So, when an agent wants to uh read a codebase, it must go through GitLab.

[00:14:30]when an agent has to write to a code base, it must go through GitLab.

[00:14:34]>> Okay.

[00:14:35]>> Right. And because we sit in the middle as the gatekeeper, uh that's why we are able to provide that level of control.

[00:14:42]>> H So GitLab is the is the gatekeeper and really the guard rail.

[00:14:46]>> Exactly. That's right.

[00:14:48]>> Okay. So I think it's become a real issue in uh in develop the development these days in terms of the cost of the platform. It's not the way it used to cost. Obviously there there's there's far more code being generated so a lot of the platforms need to raise their costs. How is GitLab addressing that issue?

[00:15:05]>> Yeah, you know there's two parts of cost for our customers. The first is the cost of the platform which is very predictable. It's based on the number of developers they have. It's charge per seat and that's always going to be there and it's part of the agentic platform and it's uh it's always going to be consistent.

[00:15:25]>> Okay. The part that is a little bit more uh concerning for customers is the cost of AI >> right because it's variable it changes uh you know the more you use AI the more um customers have to pay for the underlying tokens the underlying uh capabilities and that creates a variable amount which is hard to predict and hard to control >> right >> so uh we we're doing two things to solve that problem first we have introduced cost controls for AI when customers are using AI the dual agent platform where they can specify how many credits each team the entire company or even each person can use so they can keep keep track of it.

[00:16:10]>> Mhm. But more importantly, uh, on June 10th, we're introducing the flex buying program, which allows our customers to, uh, change how much they spend out of their total commitment towards seeds or AI or anything else they buy from GitLab.

[00:16:29]>> Okay.

[00:16:30]>> Now, that's really useful when our customers have some level of uncertainty on how many seats they need. You know, some of our customers use contractors in some on some days and and not on others.

[00:16:42]>> Uh some of our customers have uh busier periods in the year where they're going to use more AI to get their projects done faster.

[00:16:50]>> Mhm. Plex helps our customers accommodate for both where they commit to $1 value for the year and they can change every month how much is allocated to seats or how much is allocated to AI and they can always leave a portion unallocated for those busier months and take advantage of their budget when they need it the most.

[00:17:12]>> So it's going to scale as they need it.

[00:17:15]>> This will scale as they need it.

[00:17:16]Exactly.

[00:17:17]>> Gotcha. Okay. Makes a lot of sense.

[00:17:19]Well, you know, with all these new developments, what do you see going forward? Say say 3 to 5 years ahead. You know, a lot of big changes are coming up, I'm sure. What what do you think is the one thing that will surprise people the most about how software is built in this AI agent world we're going into?

[00:17:34]>> Yeah. Well, you know, maybe let me start with saying that it's clear that the software development life cycle is compressing.

[00:17:41]>> No doubt. Yeah. We're going from a place of manual software development which took weeks to go from idea to devel to production >> to agentic assist where it takes days to go from idea to production >> to autonomous software development where it'll take minutes to go to production >> and most enterprises will end up with all three modes of software development at the same time where uh GitLab our focus is to be the one platform that is the con that's the constant to support all three modes.

[00:18:14]>> Okay, >> some more some new projects, green field projects in autonomous mode, sensitive projects in manual mode and everything else in the middle.

[00:18:24]Uh and I think that's the world that mixed mode is the world that we're going to be in uh in the future especially in the enterprise. The other part of this that I think is important to point out is uh there's been a lot of talk about the role of the software engineer or the role of the developer and if that's still needed.

[00:18:44]>> Definitely. Yes. A lot of concern about that.

[00:18:46]>> Lot of concern about that. It is absolutely true that the role of a developer and a software engineer is changing.

[00:18:54]>> Mhm.

[00:18:55]>> Right. because a lot of the code writing and a lot of the tasks are being done with the help of agents and we can go a lot faster, right? But I only see the need for developers and software engineers grow >> because now there are even more architecture decisions to be made. There are even more design decisions to be made. There are even more uh customer value decisions to be made than ever before.

[00:19:21]>> Mhm.

[00:19:22]>> Right. In fact, I look at myself, look at my team, I look at everybody around me. We have never been busier.

[00:19:29]>> Really? Okay. You need the developers more than ever.

[00:19:32]>> More than ever. And and I I'll point out one more thing. Time and time again, we have proven that the human ambition rises above the capacity of us being able to get getting things done.

[00:19:45]>> Mhm.

[00:19:46]>> Right. And AI is no different. AI is giving us more capacity to do software engineering faster, better and more of it including many other things. Uh but our capac but our ambition is only growing. The backlog of the world's software is still larger than our capability to to fulfill that backlog >> and my backlog is growing faster than my capacity is growing. That's why I'm confident the number of engineers will only grow. Uh yes, there will be a difference in in what we do and there is a releving of what uh how we go about our process.

[00:20:28]>> Mhm.

[00:20:29]>> But it's only to the need will never go away.

[00:20:32]>> Mhm.

[00:20:32]>> Well, it's almost like the developer becomes a a meta developer. Not not writing code exactly but making upper level decisions. Um of course then I this some people might ask what what about the the junior developer? how can they get into the industry then?

[00:20:46]>> Yeah. So, uh I think it's um it's the same thing. I think it's we need people to orchestrate their agents, right? When uh when we don't orchestrate agents, when there is no human in the loop.

[00:21:03]>> Mhm.

[00:21:04]>> That's when we end up with the chaos that we've been observing. Right.

[00:21:08]>> Chaos of reliability, chaos of security, chaos of cost, chaos of um uh artificial confidence.

[00:21:17]>> Mhm.

[00:21:18]>> Right. Uh and you know the control that we provide with our platform, it's for humans, senior developers, junior developers to to to enforce that so we can get agentic speed and control at the same time.

[00:21:32]>> I like it. U Monov, I think you said it.

[00:21:34]It's a lot of good stuff. I learned a lot. Uh I really appreciate you sharing expertise today.

[00:21:38]>> Of course. Absolutely. Great to be here.