What Happens When Coding Agents See Your Entire System

SPEAKER_1 (0:00)

The best part of Waking Up? A full cup of Folger's Coffee and music on Full Blast.

Shop Folgers Instant Coffee and more at your nearest retailer.

SPEAKER_2 (0:31)

This episode is brought to you by Google Chrome. You think you know a browser, but Gemini and Chrome, that's new. It can help you with practically anything on the web, like restoring a vintage motorcycle from a 50 page restoration blog, or finally break down that long article you've had open for weeks. Gemini and Chrome is here for it. Ready to make anything online make sense? There's no place like Chrome. Check responses set up require compatibility and availability varies 18 plus.

Mat Jones (0:56)

It's applied it into test, great, this is working. Applied it into live, everything collapsed. Naturally then I panicked.

Patrick Akil (1:03)

This is Matt Jones, Staff Engineer at Adyen, and he built the thing every engineer wants right now.

Mat Jones (1:09)

All of a sudden we pointed it out live, seven million spans a second. We can actually deliver this, and it did become bigger than what we thought it was going to be.

Patrick Akil (1:16)

What happens when agents can see every single system you run, and trace that back to the code?

Mat Jones (1:21)

As humans, we can see a lot more, and then agents can see way more. So we thought, you know what? We're just going to turn it off. We're just going to not check.

Patrick Akil (1:31)

Are you happy with that choice down the line? Is there a single thing that you would say, this has really made me more productive either with agents or just in general in building software?

Mat Jones (1:46)

I think the fact that you can now multitask a lot heavier. So my workflow at the moment is, Mondays especially, work from home, it's great. So I will open DevIn Factory, whatever, and then I will have something like four or five agents next to each other. Like two monitors, agents, agents, agents, agents, and I'm having conversation with one for OpenSpecExplore, and then a conversation with another one for OpenSpecExplore, and it's like there's all of these different things that are kind of, okay, we need to start figuring out. And then as soon as we're good and like the spec is built, I just go, okay, cool, we'll apply. And then apply another one, and it's like I've got my workflow set up, so it will create a new Git work tree, and it will move everything over there. And then we've also built this internal tool called Atrium to allow agents to communicate.

So once something starts a bit of work, it logs into Atrium, it's kind of like a chat app, kind of memory app.

And then they will kind of say, hey, I'm gonna start working on this, and here's the files that I'm gonna be working on. So then when another agent starts applying the work as well, they move to a Git work tree, they register with Atrium, and then they'll see, oh, actually, there's another agent who's touching the files that they do, and then they message each other, which is great. And then sometimes you'll see one holding off whilst the other one is changing the file, and then it'll go, oh, I've changed this file, here's my changes, and then the other one will go, cool, do we actually, like, are we going to have conflict? Oh, we can carry on.

So it's great. So it's like, we've kind of built it out so we can have multiple agents running at once, and then have those agents not step on each other's toes, and then it's just kind of off you go. So it's like the ability to then go five, six agents at the same time, ah, game changer, because it just means it's features, features, features, features, features, it's great.

Patrick Akil (3:37)

And Atrium was built within Adyen specifically to solve these kind of parallel problems.

Mat Jones (3:42)

Yeah, Atrium was kind of built as a little bit of a side project, because I originally was starting multiple agents at the same time. And the good thing with Architect is because it's kind of away from our main monorebo, a little bit more freedom there.

So the fact that I was using so many agents at the same time in order to quickly build out Architect, before I was really integrating work trees on them, I would see them really kind of conflicts everywhere. And then they would try and commit something and they go, oh, there's other things in the work tree that's been modified. I don't really know what to do here. And they're getting all confused. So I was just like, okay, they need to be able to talk to each other.