Jake Villas on Autonomous Coding Agents for Aerospace & Defense

Jeff Kavanaugh:

I'm Jeff Kavanaugh, head of the Infosys Knowledge Institute. We're here at the Mission Ready event in Melbourne, Florida, and I'm joined by Jake Villas, VP of Industrials at Cognition, the company behind Devin, one of the most talked about AI agents in enterprise software today. Jake, welcome.

Jake Villas:

Thank you for having me.

Jeff Kavanaugh:

When Cognition calls Devin an autonomous software engineer rather than a coding assistant. What does that mean?

Jake Villas:

Yeah, there's quite a few coding assistants in the market today that go and help people be 30%, 40% more productive. But there is still a lot of guiding that takes place when working with those agents, which are often coined as coding assistants. Devin is able to take a task from planning and take it all the way through to a PR. Because Devin has, it spins up its own microVM, can do all the testing, linting, writing of the code. If something happens where, you know, let's say it goes and writes code and it doesn't pass CI/CD, it'll actually see that and do a self-validation, self-healing loop to actually fix that so that it does pass CI/CD. And the end result with Devin is always a PR. And so there's a pretty big difference in terms of the amount of guidance the agent needs when someone is using Devin versus an AI coding assistant because it could go through that entire process and create a PR.

Jeff Kavanaugh:

Devin, Cognition, went literally zero to 500 million in 18 months. What's driving it? It's insane. And who are the major customers? And More importantly, why has this taken off?

Jake Villas:

Yeah. I think when you look at the problems that we're solving, a lot of these companies have a massive backlog of work that needs to be done. Often organizations have probably 30 to 40% of their engineers are doing things just to keep the lights on. just to keep the lights on. Another 30 to 40% is spent on modernizing their tech stack so that they can build. The technical debt remediation. And the rest is actually net new features that's going to bring value to their customers. So Devin is able to go into these really complex organizations that are highly regulated. Our first customers were Citi and Goldman Sachs. And understand that code base, the architecture that their enterprise has been built on, and go accelerate those modernizations. Automate a lot of the security vulnerabilities that takes time away from these developers, automate a bunch of the service ticket triage requests that come in. So Devin is actually going and it's really creating a whole new level of capacity for these software engineering teams that didn't exist before.

Jeff Kavanaugh:

So you weren't shut out by being in a regulated industry. You actually worked within those guardrails?

Jake Villas:

Yeah. Worked within those guardrails. A lot of these different organizations some of them even have their own programming languages, and Devin is able to go and understand that and work with that and do any sort of transformation that needs to take place, or just use that information, that knowledge, as developers are going and doing work and making sure it fits within the standards and best practices that are embedded inside of these organizations.

Jeff Kavanaugh:

You mentioning Devin in each case here, which obviously you're with the company, is the solution because of the class of software, the application that you have, or is it something very distinct about Devin itself?

Jake Villas:

So Devin is a compound AI system. And we are doing a bunch of different orchestration and using different models underneath the hood of Devin. And so, we may have a customer go and do a simple Java upgrade. We've done quite a few of these and have a playbook of what is the best way to actually go do that, what's going to be most effective. And so we have playbooks for those types of things. But beyond just the models and the orchestration, the architecture that Devin is built on is a form factor difference from a lot of different solutions out there today. Devin spins up its own microVM in the cloud. Each Devin session that runs actually has its own web shell, browser, and IDE that allows Devin to go and do this work in an autonomous way.

Jeff Kavanaugh:

And secure way, evidently, because it's its own VM.

Jake Villas:

And you can validate and verify everything that Devin does.

Jeff Kavanaugh:

That's probably the most important thing you said, besides the guardrails pieces, you can validate it. It's not a black box.

Jake Villas:

Exactly. And I work with a lot of the largest automotive companies in the world, aerospace and defense companies, and you think about the space that they are working in, mistakes are a huge issue. You can't have mistakes. If I'm working with a retail customer, they have a mistake on their website, they might lose revenue. When we're talking about aerospace and we're talking about automotive, those industries, mistakes can be fatal. And so working within those guardrails and passing CI/CD, doing all the different things that are extremely important to enable the organization to run safely and effectively is so important. That's why we started with complex, highly rated businesses because we really wanted to make sure we are solving the biggest and most difficult problems. And because they're so rule-oriented, they're used to it, and they actually embrace that.

Jeff Kavanaugh:

Walk me through the first 90 days or so when someone adopts Devin, and maybe how that's different than a traditional system.

Jake Villas:

When we go and deploy in an organization, we'll build out playbooks where we'll run sessions and figure out what's the right way to do this specific project. And after we've done that successfully, think about a migration might be thousands of processes that are pretty similar. And so we will build out these playbooks.

Jeff Kavanaugh:

And you're incorporating all those learnings along the way.

Jake Villas:

We always talk about the first day of using Devin, it's the dumbest that Devin's ever going to be, because Devin is constantly learning the environment of the customers that we're working with. And once you go and do some of these things once, you have a repeatable process to go from zero to one to now one to a 1,000 and really accelerate the work that the team is doing. And I think part of that is we're now enabling these engineers to go from bricklayers to architects and managers of fleets of agents, where they're really focused on the craftsmanship of what are they trying to build.

Jeff Kavanaugh:

Your company talks about the difference between not having a coding assistant, but a true architect. Walk me through that.

Jake Villas:

Well, working with a coding assistant, you're working hand-in-hand with an agent, and you're still really working on specific units of work. But with Devin, you are now working with an army of engineers, and the process becomes, what do I want to build? I have the resources to go build it. And so people are really focused much more on the craft of what they're trying to build and using agents and being able to manage them effectively while they're doing this complex work that enables them to really go and scale at a speed that's quite a bit different from just working with a coding assistant.

Jeff Kavanaugh:

For an aerospace and defense organization, things like code quality, auditability, security clearances, you've got to have them, non-negotiable. What does responsible deployment of an autonomous coding agent look like?

Jake Villas:

Yeah. Firstly, the way that we deploy is critical. So our customers in our aerospace and defense segment are working inside of our FedRAMP high environment that is purpose-built for government agencies and the most highly classified information. So that's step one. Step two, when you think about just how Devin actually goes and does work, it works inside of its own self-validation where it's looking at all the test coverage that the code has. It's doing a validation and verification within any code that it writes, and it's making sure that it passes CI/CD. So when you think about the best practices an organization has, we're using that data specifically and that code specifically to ensure that anything that Devin does is following those best practices. The companies can input their knowledge within their organization and make sure that Devin is following those protocols exactly as they're expecting it to be. And actually, we often find if you think about being able to embed that inside of your SDLC with an agent, developers at large organizations, you've got standards, but they may or may not always follow these standards. If you have an agent say, this is exactly how we want this done, you have a sense of here are the guardrails that you need to work in, how things should be done. So Devin really allows for and enables those best practices and protocols that are critical inside of the aerospace and defense industry.

Jeff Kavanaugh:

You've got a wide background with consulting and working with a lot of different large organizations. From your experience, what do large companies get wrong when they're trying to adopt agentic AI? So what is the one thing that the people you work with or any large company leaders can do to think differently about agentic AI and maybe how it can be adopted into their company?

Jake Villas:

I would say a lot of organizations are just trying to do AI instead of thinking about what are the business problems that we want to solve with AI? How do we actually want to change our workforce with AI? Because when you bring AI into an organization, if they're doing things correctly, roles should change.

Jeff Kavanaugh:

Last question. This event today was called Mission Ready AI. Mission ready implies something proven, hardened, and trusted. How close is agentic AI to being mission ready for A&D, aerospace and defense today?

Jake Villas:

Yeah. I think we're already there. It's already being deployed in so many of these aerospace and defense companies that I think we are there for anything in the commercial space. In the classified space, we are working with a number of different organizations helping them from a software engineering perspective. So we're already there. I think the gains that we're seeing in these organizations is pretty substantial because from a tooling perspective, many of them have always been hamstrung by what they're actually allowed to use because tools are not FedRAMP high certified. You just simply can't use them. And I think now with tooling being certified from a FedRAMP high perspective and actually doing work inside these organizations, we're there. Companies are using it today and seeing a lot of results as a result of using it.

Jeff Kavanaugh:

Well, that's exciting. Jake, thank you for sharing your insights on where agentic AI is headed and what it means for enterprise, especially A&D putting it to use.

Jake Villas:

My pleasure.

Jeff Kavanaugh:

I'm Jeff Kavanaugh. Until next time, keep learning and keep sharing.