Changing Software Development with Frontier Agents

Bloomberg

Watch Changing Software Development with Frontier Agents - Bloomberg on Dailymotion

Transcript

00:00Considering where AWS sits in the market and everyone that attends reInvent, probably the

00:05question most people would have is, if I'm a large enterprise and I have multiple teams,

00:10multiple desks doing multiple things, how do I deploy three of these frontier agents?

00:15Yeah. Small, let me start out with saying, here at AWS, we are building the foundation

00:21for billions of agents, and that's an understatement. So what we launched with these

00:26frontier agents is essentially a new category of agents because while so much of activity is

00:33happening in agents, they are historically being more assistance to humans. And while we deployed

00:40it even within Amazon across tens of thousands of developers to help with software development,

00:45we were actually wanting to push the envelope and being able to actually do a lot more where they

00:52are actually autonomous, where you don't need humans to constantly steer them. And they're massively

00:57scalable and they can run longer. To be autonomous, they need to have an end goal.

01:03That's right. How does that work? So it's a case of saying, by the way, you used the term AI

01:08assistant. Yeah. What we're now really talking about is AI coworker. That's right. And that is the

01:13key thing here. It's just like what these frontier agents are doing is completely transforming how

01:18software is getting done in these teams. So much activity happens primarily in like software

01:25development, but now these agentic teammates, they show up as another teammate for software development.

01:32They pull up actually jobs from like GitHub task and start coding and clearing backlog. And if you're in

01:40the middle of the night actually getting page to deal with an issue happening with your website,

01:46this agent first takes the first page and says, oh, let me look into what's happening and unpack and

01:51then find out the root cause so that you can actually quickly resolve it and go back to bed. Like I've

01:56been on call and I know I could use that help and be proactive so that those issues don't happen. And

02:03third is actually a security teammate. Right. So it shows up again. And most of security has always been about

02:09afterthought. One of the things we are changing the game here is actually meeting developers, even

02:15before they write a single line of code, when they're designing software, we catch them right there and say,

02:21here are the things you should be aware of. And when they write code, we say, oh, don't do that. Do this and

02:26actually do penetration testing all before shipping. So now you can view it as like you have the best in class developer,

02:35ops and security engineers showing up as virtual teammates for every software. You say best in class. What evidence does

02:42AWS have that in real world deployments running these three agents in parallel autonomously

02:49has a clear benefit, either efficiency, productivity? What data can you share?

02:53Yeah, I'll give a couple of examples. One is in the DevOps agent that we launched, we ran it across

03:00thousands of escalations that happened this year alone on like incidents and so forth. These agents

03:08actually correctly identified the root cause 86% of time. Right. That is like really impressive.

03:1486, 86%. 86%. And the second one is Commonwealth Bank of Australia. They actually have a huge cloud

03:22infrastructure running across 1700 accounts. And they put this DevOps agent test to test and

03:30for a very complex networking stack to debug it. And what they said, what it would take them like

03:36hours to debug? This was able to do it in minutes. And these are just few simple examples. Same thing

03:43what's going on with security agents. Smugmug is a great photo company and I'm a big customer. And they

03:51are completely automating the security operations altogether using AWS security agent. And these are

03:59all the beginning of how this is going to change the game in a big way. Because what you want these

04:05agenting teammates to be is always be there and you delegate some of the boring drudgery associated with

04:12like ops and security and backlog on development so that developers are empowered to be extremely

04:19creative. And you will start seeing 5 to 10x improvement in productivity in a big way.

04:25To recap, it's three agents. Kiro is software development. You have the AWS security agent,

04:30the AWS DevOps agent. They are designed to run for hours autonomously. To some, that's disconcerting.

04:37Right. And I go back to my other question on, you know, how much concern is there? In the last segment,

04:43Killeen talked to me about observability, understanding in real time what's doing. But that sounds to me like

04:50it's human supervision. It's a great question. I mean, this is the nuance and the art of balance you have to do.

04:57One, you want these agents to be autonomous and massively scalable, but you also want them to

05:04not go off the rails. So first, we are built it with the right guardrails. And second, even as part

05:09of development workflow, these agents can go actually take a goal from a developer to say,

05:15hey, go work on upgrading this piece of code to the latest SDK and then start upgrading. But then as a

05:22final step, it might send it for a code review to the human and say, take a look and make sure I'm

05:27okay. And once they say, I'm okay, I think it's fine. Or they can even configure it to say, if it

05:32passes all these tests, you're good to ship it so that you have built in guardrails either through

05:37human or automated tests. And these are the kind of things now we are able to actually innovate with

05:44the customers as well. And even there, we have done some pure innovation like introduce automated

05:51mathematically provable property based testing. So that means now developers can specify their goal

05:59and we can mathematically prove with things like Kiro that whatever it is generating can be tested

06:05exactly that way. And it's accurate. So that is like kind of the game changer.

Category

Transcript

Be the first to comment

Recommended