How Capital One supports 14,000 technologists with one pipeline | Ameesh Paleja

**Andrew Zigler** (0:06)
Welcome to Dev Interrupted. I'm your host, Andrew Zigler.

**Ben Lloyd Pearson** (0:09)
And I'm your host, Ben Lloyd Pearson. This week, we welcome Capital One's Ameesh Paleja, EVP of Enterprise Platforms, to discuss how standardization serves as the unsung hero that unlocks scale for his fleet of 14,000 technologists. He explains his strategy for giving engineers an Iron Man suit of AI tools to eliminate mundane work, all while maintaining the strict resilience required in banking. But first, let's cover this week's news with ChatGPT fumbling breaking news, Inside Big Tech's Knife Fight, and Murder Your Friday Deploy Freeze. Andrew, what do you want to cover first?

**Andrew Zigler** (0:47)
Okay, well, that sounds like a great lineup of new stuff, but it's the top of the year, Ben. It's January. I'm excited to kick off 2026 So everyone's talking about predictions right now. I want to talk about some of the fun AI predictions that have come across my desk, at least, that I'm sure others have seen as well. And there's one particular article that I'd love to talk with you about. So let's do it. First off, I think that this year is going to be framed by big tech capital expenditures and AI infrastructure exceeding numbers never before seen. And they were already eye watering and staggering to begin with. We're talking about projects that are going to exceed 500 billion in 2026, up from 400 billion last year, with companies like Google and Microsoft and Amazon, obviously leading the way.
And as AI model capabilities continue to progress, the predictions this year seem to revolve around models being able to perform long tail tasks with higher and higher accuracy, beginning to closer mirror the success rate of entry level engineers even for a certain task like picking up beginner JIRA tickets. But the most important and interesting takeaway for me was about how I think we're entering the limitations of the context window expanding. And this year is going to be the year of us doubling down what actually goes into those one million or so tokens that we're going to use to build the future of engineering and context engineering is going to continue to take a front row seat. But I think that most of all, 2026, it won't be the year that AI just takes off and just is uninterrupted. It'll be instead the year that AI gets procured. You're going to see it more formally ironed and cemented into large companies and be formalized in the processes that we haven't seen it in before. It's going to go from cool demo to line item on the budget. I think this is the year for that. Ben, what do you think about some of these predictions?

**Ben Lloyd Pearson** (2:40)
Yeah. We'll be sharing a link in the show notes to this sub stack article. It comes from Understanding AI. They brought together a bunch of experts from the field to provide a bunch of predictions for 2026
Some really great ones, many that I agree with, some that I disagree with. I think the one that stuck out to me the most was about how context windows will likely stay around the 1 million token mark. I kind of agree with this one simply because we have a lot more to gain right now from applying these models to new situations rather than trying to make the models themselves more comprehensive. I don't actually need a 1 million token context window very often, but when I do, they work very well and I've yet to encounter a text problem big enough that it can't be handled by a model that's that big. I feel like if you were to need something that exceeded a million tokens, you can probably optimize the data you're inputting into it rather than trying to just add a bigger context window. And then, yeah, you alluded to these predictions that models will be able to solve longer-term problems, like the idea of how frequently can they solve a problem that would take a human about five hours to accomplish, and today we seem to be at about 50% of the time the frontier models can successfully accomplish something that might take a human five hours to accomplish. And it's interesting to think that that might become 20 hours over the next year. That would be a pretty substantial improvement. But one thing that stuck out to me that I wanted to get your opinion on, Andrew, is the prediction they had about MCP and how it would maybe become irrelevant this year. What do you think about that?

**Andrew Zigler** (4:23)
I definitely think there might be some accuracy with that, not just about MCP specifically, but just about just like bubbling up of protocols that we're seeing kind of universally. We're trying to solve a lot of fuzzy problems right now working with agents and it's getting tackled in a lot of ways. There's obvious ones like the spec and the plan and how you actually direct your plan. But there's also harder to define ones like session-to-session memory and persistence, and even simple working processes like having Git work trees and non-collisions when trying to orchestrate these agents at scale and in unison. So MCP, when it came onto the scene as a protocol, tools were starting to become a thing that agents and LLMs were picking up and using.