Brex’s AI Hail Mary — With CTO James Reggio (acquired for $5B by Capital One!)

We have, like, three pillars for our AI strategy. We have our corporate AI strategy, which is how are we going to adopt and, like, buy AI tooling, um, across the business in basically every single function to be able to 10X, uh, our workflows.

Then we have our operational AI strategy, which is how are we going to buy and build, uh, solutions that enable us to lower our cost of operations as a financial institution. And then the final pillar is the product AI pillar, which is, like, how are we going to introduce new features, uh, that, um, enable Brex to be a part of the corporate AI pillar of our customers?

It's like we wanna build features and be a, a, a solution that somebody else is saying to their board, "Hey, we, we adopted Brex, and this is part of our corporate AI strategy."

Hey, everyone. Welcome to the Latent Space Podcast. This is Alessio, founder of Kernel Labs, and I'm joined by Swyx, editor of Latent Space.

Hey, hey, hey, and we're here with James Reggio, CTO at Brex. Welcome.

Hey, thank you for having me.

Thanks for visiting, uh, from up in Seattle, where, uh, I, I've been a little bit. It's cold up there, huh?

Yeah, and we have an atmospheric river hitting the, the, the city right now, so a lot of it-

A lot of it flowing?

Yeah. Well, yeah, it's, uh... We're getting, we're getting the full-on winter effect right now.

Well, you're, you're here to-- We talk about the sort of AI transformation within Brex. There's a lot of, um, interesting tidbits that we're gonna draw from your article, but also your background. You have got a wide array of experience from Stripe to, um, Banter to Convoy.

Mm-hmm.

And, uh, I think also mostly I'm interested in your journey as, as one of the rare people that have transitioned from, like, a mobile engineering leader to a CTO, which I think is also a bit more rare. I, I used to have this comment in the past where there's a career ceiling for people who work on client-only things-

Mm-hmm

... where usually they don't hit CTO, whereas they typically promote the, the back-end people or the back-end cloud infra people to CTO.

Yeah, you know, it's, it's something that I, I hear, hear fairly, fairly frequently because, um, there aren't that many folks with a front-end background who reach this level of leadership, and it's exciting for me to be able to represent that group.

But I... I'll say that even though my resume kinda reflects that I've been more on the, the front end of things, i- it's probably more my experience as a founder a couple times over that actually helped me get to this, this level of my career working for somebody else.

Becoming CTO is very much like a leadership and, and, like, general business role as much as it is a technical role, and so I think it was more the skills that I built from starting companies and, and trying to build those up made me a decent fit and enabled me to get the nod from, from Pedro to take this on as my predecessor left about two years ago.

Yeah. One thing, I'm curious your guys' commentary. This is a little bit broad-

Mm-hmm

... unscheduled, but a lot of startups are bragging about how many ex-founders they have. And yes, to some extent you want people with the founder mentality and agency, which is-

Mm-hmm

... what, what you did, to be your employees and to, to take initiative in the company. But also I wonder if it's becoming anti-signal sometimes. I don't know if you've thought about this.

I think it's more about the churn for me, especially when people are hiring ex-founders. It's like if you're truly of the founder gene, it's kinda hard to just stay somewhere as like an IC for too long.

Ah.

And then it's like, all right, I joined this thing, and then in one year I'm back to being a founder. I'm curious for you.

Yeah.

What was your... I'm sure you thought about leaving and, like, doing another company and say-

In fact, that was, that was the, the alternative I was considering even at the time that I got the phone call where they made me the offer to become CTO. I was thinking about leaving to go start a company.

And, uh, you know, I think what's interesting about it, we, we actually launched, uh, sort of like a new recruiting and employee value proposition for Brex a couple months ago called Quitters Welcome, where we actually intentionally are leaning into this idea that, uh, we have a disproportionate number of folks who go on to become founders or, like, heads of a department when they leave our company, and, and we celebrate that.

It's actually something that, uh, I, I'm very proud of. And that means that, like, we, um, we welcome in people who want to get a different experience. I think that there's certainly, like, a, a lot of founders who don't make it, uh, don't scale their own businesses to the s- to the scale that we've achieved at Brex, so there's something to be learned when they come in.

And then we're very happy to, like, uh, support people on their way out. And so I, I actually really like hiring for- former founders or future founders. Uh, the one value proposition I find that's most relevant, 'cause a lot of the folks we're hiring as AI engineers are kinda folks that are either, like, winding down their companies or, or are considering, um, maybe running AI startup.

The, the thing that resonates the most with them is that we oftentimes can give them problems to solve that are interesting, problems that may, maybe they even want to want to, like, build their own startup around, but with instant distribution, right?

Like, that, that is the, that is the allure is it's like you can come into this business and build, uh, like, financial AI applications and instantly have it deployed to roughly forty thousand, uh, customers across, uh, you know, the Fortune one hundred down to, you know, tens of thousands of startups.

So that, that's what is, I think, appealing to founders. But, but the, uh, the challenge then is making sure that we set them up for success in, in an environment that still feels a little bit like the startup that they might build themselves versus, like, something that's too corporate.

Yeah. Instead of doing your own company and then coming to you and be like, "Can I integrate into Brex? You get all the data-

Yes. Exactly

... into my financial institution." How's the engineering team structure?

Yeah. So we have about, uh, three hundred people in engineering, like three hundred and fifty, uh, total across EPD. And, uh, for the most part, we structure around, um, our product domains, and so this means that Brex is a corporate card.

Uh, it's also a corporate bank account. It's expense management, travel, and accounting. And so we, we actually have sort of full stack, uh, product domains, uh, that are roughly, like, thirty, forty people for each of those that, um, have everything from, like, the low-level infrastructure up to the, the web and mobile experiences.

Mm-hmm.

That's generally, like, the structure of, of our, of our engineering, um, organization. And then we have, uh, naturally, like, a, an organization that focuses on infrastructure security, um, IT, uh, and then there are two, uh, additional centers of excellence that we've kinda built that kind of violate that org design, uh, where we've felt the need to, uh, to put more focus or, like, operate slightly differently, and AI is one of those areas where we have another team of just roughly about ten people, um, who are focused primar- primarily on LLM applications.

And we wanted to Create a bit of a separation there because the way that we were thinking about this, and this is actually something we did this summer, is we, we paused and asked ourselves on, on our AI journey towards like infusing our product, uh, with AI and generating customer value.

We asked ourselves like, "What would a company that was founded today to disrupt Brex look like?" And then we tried to basically use the answer to that question to form this team, uh, internally. So it's a little bit off to the side.

Ideally, everybody kind of comes up to speed and, um, contributes, uh, you know, LLM features, but, uh, but we have this sort of off on the side right now in a centralized manner.

What's the difference in AI adoption for those teams? So like are the people on the LLM team like much bigger Cursor users, Clock.co users, or like do you see similar diffusion?

It's actually fairly, fairly uniform across the entire engineering, uh, uh, department. It, it's actually kind of funny, like one of our, our largest Cursor users is actually an engineering manager. So like, and I, I think that this also just like speaks to our core value of operate at all levels, where we want all of our EMs and everybody in leadership to still basically do the job, uh, that they're managing-

Yeah

... manage the work. So it actually is... I, I, I, I think the journey of getting everybody into using agentic coding was not sort of exclusive to like the, the AI group.

Yeah. Um, I, in fact, I think this podcast was actually set up because I cold outreached to Pedro-

Mm-hmm

... because he s- he tweeted this. I, I, I assume this is the center of Brex is.

Yep.

It says, "I started a new company inside Brex to build the future of agentic finance. No BS, just builders building nine in six and pr- pushing production-grade agents to 30,000 finance teams," now 40,000. Um, and then he actually has like a little job description, which I think is really interesting.

Uh, but I'll skip that and go straight to, "Brex has accelerated growth 5X and cut burn 99% in the past 18 months." I assume that's a mix of internal AI automation and other stuff.

Mm-hmm.

Um, but we're... Basically, I wanted to put some headline numbers up front to impress people-

Yeah

... before we dig into the details.

Yeah, absolutely, and you're, you're correct. That's the, that's the team. Now we have this, uh, like AI team. We're actually-

Very young team.

What was that?

Very young team.

Yeah, it's very young. I mean, it's... and it's been really interesting. The, the composition of the team is like very young, like AI native, like 20-year-olds who basically grew up with the tech, um, kind of paired off with like more like staff level software engineers that have been at Brex for a little while who can kinda navigate like the existing code bases and like understand, uh, the product and the customer deeply.

Like we've formed these really, uh, couple of tight, tight-knit pods in the AI org where it's like three people. Generally, somebody who has like more of a product, a customer-focused background, that like staff engineer who knows, uh, where the skeletons are, and then like, uh, a much younger like AI native engineer who, um, can just do things with, uh, with agents that like the rest of us, uh, dinosaurs maybe, uh, don't, don't...

uh, can't either dream of or like or where are... I, I think, I think part of it is like sometimes the too much experience or too much knowledge of how to solve a problem can actually be an impediment to thinking, uh, differently about it and thinking about it from like an AI first lens.

But yes, we, we've been, we've been slowly growing that team just in the same way that like a pre-seed startup, you wanna be very, very careful about talent density and like very deliberate, like only hire when you absolutely need it.

And so, yeah, at this point, it's just about 10 people, and I think it was probably four or five people. Uh, I think everybody was actually in the photo that was attached to that tweet, uh, when Pedro put that out a couple months ago.

Yeah, we'll put it up. Yeah, that's a photo at 1:20 a.m. in a, on a Friday.

Yes. Oh, yeah, yeah. 'Cause we'd, we always do, uh, we always do like Friday, Friday demos and, and like that's a time for everybody to get like kind of exec review time. And so, uh-

Everyone's in Seattle?

Um, those folks were all in Seattle. Uh, but w- they're actually geo- geographically distributed. Uh, we have s- a couple folks here, a couple in São Paulo, a couple in Seattle.

Hmm. How... At Decibel, we have this like AI center of excellence-

Mm-hmm

... which are basically the people running these teams across companies.

Yep.

How do you make the other engineers not feel like you're not special? I think that's something that I hear-

Yeah

... a lot is like, "Hey, you know, why aren't these people working all the cool LLM things, and like I'm stuck working on, you know, the KYC integration with whatever?"

Yeah.

You know what I mean? It's like, how do you build that culture?

You know, it's interesting. I, I thought that that would be more of a problem, but the benefit of having really optimized our engineering culture around business impact actually causes it to cut in the other direction, where for folks, some folks don't want to work on the AI products because it doesn't have as much clear, direct like business impact right now.

Doesn't, doesn't impact revenues directly. And so I, uh, I think folks for the most part, uh, we've, we've enabled folks who have a strong desire to work on, on, um, AI products to, to join that team. Like somebody, somebody transferred out of our expense management organization to come over there because they're really passionate about taking like their knowledge of like policy evaluation and, and bringing it into the, the AI, uh, uh, team.

But for the most part, I think everybody understands like how their work, uh, ladders up, and maybe there's some like friendly li- rivalry because like the folks who, say, work on a card product, they, they drive 60% of our direct revenue, and so they, uh, they're pretty happy with that, and, uh, and they don't feel like they're being left out.

Uh, and I will also say, um, as you probably saw in, in this, this piece that we, we, uh, put out with, uh, First Round, there is a lot of smaller applications of LLMs peppered throughout all of our product and, and operations teams.

It's just some of the more novel like agentic layer that sits on top of Brex that has been put together like in this, in this sort of isolated team. So it's not like folks aren't getting to, to build with LLMs or use LLMs on a daily basis.

Yeah, maybe run people through the Brex agent platform. We'll put the diagram in the video where you had the LLM gateway. You have like the whole MCP layer. We just had David, the creator of MCP, right before you.

Cool.

So this is very timely.

Yeah.

Um, yeah, how did you start building that? What's the architecture?

Yeah, the architecture, you know, I, I think simple is, uh, is elegant, and we, we've had basically an LLM gateway and, and, uh, a basic canned rolled platform, uh, from the very early days. In fact, right before being tapped to become CTO, I was leading, uh, like a AI, uh, labs team internally, uh, in the wake of like the announcement of ChatGPT.

You know, everybody saw this new technology and said, "Hey, what are we gonna do with it?" And so one of the first things that we did, um, I think January 2023, that would've been Uh, was try to put together some internal infrastructure that made it possible for us to deploy pro-- deploy, manage, version, and eval prompts, uh, and then be able to manage, uh, like data egress and model routing and, uh, have some very basic like observability and cost monitoring, uh, in an LLM gateway.

So that's, that's infrastructure that we stood up, and it still continues to power a lot of those smaller, uh, more, let's say like precise applications of LLMs. So like for instance, we've, uh, we've set up a completely automated, uh, pipeline for, um, evaluating, uh, customer applications to get them onboarded instantly to Brex, which is something that used to require, um, human intervention either for underwriting or KYC.

But now we basically have a series of, of agents and, um, and particularly like research agents that will go and do the work that humans would normally do. And so that's running on top of this, uh, this hand-rolled, uh, framework.

And then for the agents on Brex that we announced in our fall release, which is like this agentic layer that we're building that sort of sits on top of Brex and can embody workflows that a finance team would normally, uh, hire humans for.

We've actually, uh, started using Mastra for that as like the kind of primary-

Ooh.

Primary framework for, for accelerating this. We actually have built everything in TypeScript, um, uh, which is another like technology choice that's, uh, that answers the question of like, what would we do if we started Brex today, but isn't the case for all of our existing backend code, which is either Kotlin or Elixir.

And then we have, uh, we have a mix of pgvector, Pinecone, and, uh, like I think what we've seen is we're always, we're always reevaluating the tech and framework choices as we go, uh, because the half-life of code has declined so significantly with agentic coding.

It's actually quite, uh, easy for us and for anyone else to, to kind of try on for size a variety of different pieces of tech to, to figure out what is going to be most ergonomic for solving the problem.

Double click on Mastra. That's a new choice, an interesting one.

Yeah. I mean, I think that the main, the main reason that w-we adopted Mastra was that it provided the ergonomics that we were actually, uh, that... The ergonomics of Mastra are quite similar to the internal, um, LLM framework that we built two and a half years ago.

Uh, whereas like LangChain was available at the time two and a half, three years ago, uh, it didn't quite feel, uh, right to us when we were trying to, um... It, it kind of addressed the things that weren't the, the pieces that we, we needed to address, which was like being able to have really simple, um, observability and, and, um, logging, tracing, um-

LangChain didn't do it?

It-- I mean, at that time it didn't. I think it was really, I think it was-

Well, they fixed that.

Yeah. No, they certainly-

Yeah.

They certainly did. But, but, but so like we, we did... I'm trying to remember 'cause this is now ancient history. Uh, we evaluated LangChain, turned off of it, built our own thing, and then as we were looking, we, we kind of want to deprecate this internal framework that we built because at the end of the day, it's not leveraged for us to maintain that.

Uh, and Mastra ended up fitting the bill for, um, for the, the feature set that we were looking for. And I think what, what's been interesting is about half of the, the applications that we're, we're building right now, um, on the, the agent layer are running on Mastra, and then the other half are actually still running on like yet another internally developed framework-

Mm.

Which is a framework that's focused more on networks of agents, so sort of multi-agent orchestration versus more like strict, uh, like, you know, single turn or, uh, like workflows which are easier to use like either LangGraph or Mastra.

Tell us about your multi-agent framework. I mean, that's... Uh, w-what are the design considerations? Um, why, why, why is this the first we're hearing about it?

Yeah, yeah. So it's funny. A, a, a big, big reason why we haven't written more about this is that it continues to evolve quite a bit, and I, I feel like we, we actually had a blog post that we were going to put out in conjunction with the fall release, uh, talking about how we built this.

And by the time that we finished, uh, you know, the blog post and had all the package ready, it was already like halfway outdated. And so the way that this has started to emerge is this multi-agent network approach to implementation was when we were trying to scale up our, um, sort of consumer-grade Brex assistant.

So if you think about like Brex, uh, and our customers, uh, there's really like two very broad personas that we serve. We serve members of a finance team who are generally like going to be doing like in roles like accountant or controller or head of T&E.

For those folks, uh, they are, uh, going to be interacting with agents that are much more specific to their roles. But then the other broad cohort of, of users we have are like employees of companies that have deployed Brex.

So, you know, you go join a new company, that company uses Brex, you get your Brex card, and our goal for employees is for Brex to completely disappear. Like the best UI/UX for Brex is just the card. Like every single thing that you have to do in the software beyond just swiping the card is like an opportunity for AI to, uh, to eliminate some work for you.

And so what we thought was the right approach to solving th-- for that was to, um, was to embody like an executive assistant, uh, for every employee. Because I as an executive at Brex, I have an EA, and she knows enough about me.

She has access to my calendar, my email, has all the context on, uh, when I'm traveling and for what business purposes. And so she's basically able to do everything that I would be obligated to, to do in Brex, be it like booking travel or like doing expense documentation.

And so what we wanted to do is we wanted to build like that EA connected to the same data sources and see if we couldn't simulate that behavior so that, um, you know, you basically, you're interfaced to Brex's, uh, SMS and the card.

And when we started building that out, um, you know, the most naive like architecture for that would be to have, um, an agent with a variety of tools and maybe, maybe do some, some RAG to ensure that it has like appropriate context for the conversation.

But what we were finding is that, um- The wide range of different product lines that exist on Brex made it difficult for one, uh, like, agent to perform well, uh, being responsible from everything from, like, expense management to finding and booking travel, to answering policy and procurement questions.

And so that's when we started breaking down the problem, uh, and into, into a variety of sub-agents that sit behind an orchestrator. And obviously, this is, uh, something that can be implemented using LangGraph, or Master even has the notion of these as, like, network switches and data.

But what we found is that it was easier for us, um, when it came to being able to build evals for the system. Uh, we, we kinda just hit the eject button and built our own framework, which is one in which, um, we have agents that are able to, uh, to basically DM with other agents and have multi-turn conversations amongst themselves to coordinate to, um, to complete a task, to, uh, or, like, to complete an objective.

And, um, what's, what's been nice about that is it means that, like, you can have your Brex assistant as, like, one single, one single, like, point of contact, uh, between you as an employee and the Brex product. And then behind your assistant, um, if the company has, like, expense management turned on, you have that.

If they have reimbursements, there's another agent for that. If they, they have travel attached, they have agent for that. It actually also then facilitates, like, our conception here is that, um, it, you know, it's like generally like software encapsulation patterns taking...

like, sort of projected into the agent space. It also makes it easier for us to have, like, the team that owns and understands travel, like, be the ones to go and iterate on that without needing to worry about, like, regressing the total system, um, or needing, like, one team to own every single possible, um, act- action you could take as an employee.

And I'll say that, like, I'm still of the mindset that somebody will build a, a great framework a- and we may evolve, ultimately migrate to it, but... Or it might be us that we ultimately open source this, right?

Sure.

Like, but, um, but for us, like, this is, uh, this has worked out quite well in, like, lieu of, like, a couple other approaches that we, we tried along the way that just didn't perform well, which was to, you know, overload the l- the, the agent with a variety of tools or contextual, like, context switching, where we try to say, "Oh, this conversation looks like it's more about reimbursement, so let's, like, update the, the prompt with more reimbursement context."

Like, that was, that was another approach that we took that didn't perform as well as actually having a reimbursement agent that it would collaborate with.

Hmm.

What about MCPs as, like, sub-agents?

Oh, yeah.

It's another pattern.

The key thing there is that we-- There's actually a lot of value in having, like, multi-turn, um, uh, conversations from, like, the orchestrator or the assistant to, like, the sub-agent, whereas, like, you know, a tool call is basically just, like, one RPC.

And so oftentimes what will happen is, um, you know, let's say, let's say the, the, the user reaches out to their Brex assistant and says, "Hey, like, am I allowed... Like, h- how much am I allowed to expense per person for dinner tonight?

I'm taking my team out." Uh, and the, the, uh, you know, your assistant's gonna then reach out to the policy agent. Maybe the policy agent needs to know, in order to answer that question, maybe it needs to know, like, whether this was, um, uh, was, like, a customer event, a team event, or whether you're traveling.

And so it may actually send, instead of... and it can't just answer the question, so it's gonna reply back to the, the assistant and say, "Hey, you, I need you to ask, uh, this clarifying question." And so then the assistant will re- return to the user, ask the clarifying question, and then they'll basically have this sort of multi- multi-turn conversation across multiple agents, versus it just being encapsulated in, like, a single, uh, call and response tool call.

And so there are still, like, all the, all the sub-agents have a ton of tools. But I, I think of, like, the MCP and, and tool usage as being, like, the interface to all of our conventional, uh, imperative systems, not, not the, the AI space.

Yeah. That's the conversation we were having earlier, whether or not it should be an agent to agent-

Mm-hmm

... tool call as well.

Yeah.

Or, like, yeah, there should be, like, a chat back.

Exactly. Exactly. And that's the thing. It's like w- okay, and one of the ways that we actually grafted this into Astra before we, we built our own framework was to, was to make every sub-agent a tool. Uh, and then the input was just natural language, the output was natural language, and the...

If you needed to have multi- multi-turn, um-

Mm

... you would basically just put the whole, like, prior conversation in as you kept calling, calling the sub-agent as a tool. And it's just like at that point you're like, "Okay, the ergonomics are kinda... The framework, framework is fighting me on this."

It's actually helpful for us to basically conceive of it as an org chart and, uh, like, it's the agent org chart with, uh, with, um, you know, my EA is DM-ing other specialists, uh, and having brief conversations to support me as their client.

Yep.

That was a really good, uh, deep dive. Uh, thanks for indulging. I feel like you guys are not afraid to make your own tech, which I think is a competitive advantage. I really like that culture. Maybe I'll, we should go a bit breadth first as well.

Of course.

'Cause I think we also deep dive, uh, a little bit too much in, in one area. There's, um... And we'll, we'll put up the chart, but I'm also very interested in, like, the, the sort of internal agent stuff-

Mm-hmm

... the operational stuff, and just the, the general platform scope. So please feel free to just, like, go into your spiel on it.

Yeah, of course. So one of the things that I was trying to do at the beginning of the year, uh, as CTO, you know, I think it really fell to me to articulate what our AI strategy was as a business.

Yeah.

You know, every, every board of director was, you know, or every me- every member of our board was like, "Hey, what's your AI strategy?" And while we were doing a lot of-

We literally go, "He's got it."

Well, yeah, and, uh, but- Uh, yeah, and, and if I didn't, I, uh, I'd be in trouble. I think he also was counting on me, given that I was doing the AI, uh, organization before CTO to, to have-

That's true. Mm-hmm.

But, but a big part of it was, like, we, we were doing a lot with, with LLMs. Um, it was more, like, these little one-off features and, you know, hey, like, maybe mix in some su- suggestions here, or maybe do a little bit of ops automation over here.

But it wasn't, uh, it wasn't easy to, to kind of create, like, a verbal framework, uh, um, o- of all of these investments, and without that framework, then we weren't able to, like, set a, set a, a vision or a roadmap for, for investments.

So what we did at the beginning of the year is we Took everything that was going on, as well as all of our ambitions, all of the good ideas, as well as, like, the problems we were trying to tackle as a business this year, throw it all on the table and see if there were some ways to cluster it into a, a framework that made sense to the business, to our board, uh, to ourselves.

And we came up with... I, I think this is not particularly novel but has helped us quite a bit. We have, like, three pillars for our AI strategy. We have our corporate AI strategy, which is how are we going to adopt, uh, and, like, buy AI tooling, um, across the business in basically every single function to be able to 10X, uh, our workflows.

Then we have our operational AI strategy, which is how are we going to buy and build, uh, solutions that enable us to lower our cost of operations as a financial institution. Because I think it, it's fairly intuitive, like, financial institutions like ours face a lot of regulatory expectations, and there's just, like, a high ops burden for running our business.

And so it's sort of like a lot of kind of internal use cases, like being able to do, like, fraud detection, underwriting, KYC, um, be able to handle dispute automation on card transactions. Those, those types of operational investments are our ops AI pillar.

And then the final pillar is the product AI pillar, which is, like, how are we going to introduce new features, uh, that, um, enable Brex to be a part of the corporate AI pillar of our customers. It's like we wanna build features and be a, a, a solution that somebody else is saying to their board, "Hey, we, we adopted Brex, and this is part of our corporate AI strategy."

Yeah, yeah.

And so it's, it's kind of has this nice little feedback loop, and we, we basically, within the company, split, uh, you know, did a little bit of divide and conquer where, uh, folks in, um, IT and on our people team were more or less, uh, spending more of the effort driving on corporate AI, really, like, looking for, um, making the procurement decisions and, like, creating a culture of experimentation where we spotlight and incentivize people for trying, uh, to sort of improve their personal workflows using AI.

And then the, the pieces that I've been more involved in have been operational and product. And we were just talking about products here, which is, like, the agents on Brex and stuff. But I think that the operational AI investments have been some of the, the most sort of immediately impactful-

Mm-hmm

... uh, to the business because we have hundreds of people who work in our operations organization, and it's actually something that differentiates us because our, uh, CSAT and the quality of our, our support and service is very, very high, uh, something we're very proud of.

And so trying to figure out how can we automate significant portion of this and, uh, use LLMs in a way that doesn't degrade the customer experience, and then also kind of addresses, like, what is the future of the roles of the people who we already have working full-time for us.

So this is where, um, uh, Camilla, uh, our COO, who kinda co-wrote the, the piece, uh, with First Round, uh, with me, uh, she's been leaning really aggressively to help every member of the operations organization, um, start rethinking their role as being, uh, not people who kind of execute against an SOP but are people who are going to, like, build prompts, build evals, and, like, be- become more AI native in, like, the way that they do work.

And so a lot of the engineering we've done has been to enable folks, say, in, um, in fraud and risk to be able to, uh, to refine prompts and, uh, and add additional automation to their workflows.

Yeah. And this secret fourth pillar, the, the platform.

Yeah, yeah, exactly.

Yeah.

That is the, that is the thing that ties it all together exactly is, is the, is the platform. And I think what's been really nice is that, um, even though the platform is kind of a loose, uh, loose, loose t-term because it, it consists of a wide variety of technologies.

As I said, like, we haven't been too religious or dogmatic about everybody needed to be on one particular thing. What we've seen is that, um, by making a variety of sort of ergonomic, uh, options for building with LLMs available, it, like, really has made it easier for, for us to make a quick leap forward on operational AI.

Like, we-- As soon as we put our mind to it, and we said, like, "Look, no, we wanna hit a 80% automated acceptance rate for all, um, all startup and commercial businesses that apply for Brex. Like, we want a decision within 60 seconds that's fully touchless, no humans involved."

We were able to break that down and then actually build the, uh, build the agents, build the tools, um, on top of that platform really quickly, and it-- and a lot of those tools are the same tools that our product AI, uh, agents use as well.

I was pretty sold on the Conductor. I don't know if this is un-under exactly that bucket, the ConductorOne-

Mm-hmm. Oh, yeah

... uh, provisioning command. I was like, "Yep, I want that."

Yeah, that was actually... I'd love to talk about that. So that's, that's actually on the corporate side, and I think that this goes back to maybe another intuitive, but, but I'd say, like, bold decision that we made, which was that we're not going to, we're not gonna try to pick winners in the horse race between the foundational model providers or the, the agentic coding tools or, like, basically anywhere where there's, there's an active horse race.

What we do in-instead of, like, trying to pick a single solution is we will procure, like, a, a small number of seats, like multiple solutions, and then we'll give employees the ability to pick whatever one they want to use.

And so, for instance, like, we allow employees to basically go to in, in Slack and use ConductorOne to, uh, get a ChatGPT, a Claude, or a, a Gemini license. And basically you can just, like, build your own stack where you pick your, um, you pick your, like, chat, chat provider.

Uh, as a, as a dev you can pick, um, you know, between, like, Cursor, Windsurf, Claude Code credits, like, and, and you can basically craft your, your stack to your preference and easily switch between them. And what that does for us too is when we're going to...

Like, obviously we have sort of enterprise agreements in place for all of them for the sake of, like, the, you know, the, the privacy and non-training guarantees. But it's fun because when we go to renew these contracts, um, it, it-- we can basically resist the need to, like, do a wall-to-wall deployment.

We can say, "Hey, look, like, usage trends, they... Our, our employees are voting with their feet. They're voting with their dollars, and, you know, maybe, uh, maybe your tool isn't as, uh, as hot as it was a year ago."

Does it give you a dashboard of what people are choosing?

Yeah. Actually, we look at that. Um, we were looking at that as we're going into budgeting for next year. It's very interesting.

I would love, I would love to see that those... What-what's, you know, anything that's, like, really up, anything that's really down?

It's fascinating how, how different the landscape is every, every three, three months, and I think one of the, one of the interesting Challenges we had early on was, uh, getting folks to just, like, try, um, these tools, try to incorporate, like, agentic coding.

You know, and I, like, early on, I say, like, 12 to 18 months ago now, to, like, get folks to, to just take the time to try a new workflow. And now at this point, I think what we're seeing is, like, even if, you know, uh, a new model hits the scene, like, um, when, when Codex, uh, came out and everybody was like: "Oh, Codex is, is better at, at, uh, CodeGem, but it's a little bit slower."

Like, I find fewer folks are, like, kicking the tires on new things 'cause, like, the- they're just so comfortable with the ergonomics of their current workflow that, that, um, you know, uh, some folks are just like: "I'm gonna just stick with Cloud Code 'cause I know it now.

I've been working with it for, like, nine months, so I don't need to, to keep, uh, keep switching. I don't need-- I don't feel the incessant need to keep trying new things because I've, I've gotten... I'm an iPhone person, and I'm just, like, gonna stay with an iPhone, even, you know, even though there's some really sexy Android hardware out there."

Do you have one of the big numbers, like 80% of all of our code is written by AI or... But how, how do you measure it internally?

Yeah, no, not really. We, we... I mean, I, what we do is we'll, we'll measure, like, the attributions on the, the number of commits that, that have the, um, like, uh, co- co-authored with. Um, and we pull some of the stats, but I don't index that vol- like, in fact, I don't index on those at all.

I don't y- and honestly, like, I, I don't know how I can honest, like, honestly calculate that number.

Yeah, I agree.

Yeah. And so, so I, uh, the thing that, the thing that we're, we're really just... You know, we're at the point now with the, like, our AI agentic coding journey, where now we're trying to solve the second order effects of, like, a little bit too much slop, maybe a little, uh, not enough, um...

Yeah, exactly, not enough, like, rigor in code reviews. Um, we're trying to... The, the adoption is there, and now we have to figure out, like, how to mature in our usage of these tools, uh, so that we, you know, quality or, like, long-term maintainability doesn't suffer.

As well as, like, maybe one of the other fact- facets of being able to generate a lot more code more quickly is, like, the, the drift between team members as far as, like, understanding of the, the, the code that's in their services increases as, like, everybody's moving faster and more independently.

Uh, it, that is another sort of risk that we're starting to see. Like, you know, an incident response where folks don't know, they don't know a service as well as they, they used to because it's changed so much in the past couple months because everybody's moving more quickly.

Yeah.

Yeah, this has been a major topic for me this year on code-based understanding and slop because obviously it's so much easier to generate code, but then now we have to review it.

Mm-hmm.

And to some extent, you can't really fight AI with more AI. You can't just be like: "Oh, just throw, throw in the AI reviewer under AI code," and you solved it. Uh, and, and so, so you do need to just scale human attention, and, and I think that's something I've been pushing a little bit in terms of, like, well, you're, you're just gonna...

Like, every engineer is just gonna own more code.

Yep.

Period. And, uh, and, and be parachuted in and be expected to ramp up and be, be productive and also fix bugs and if you're on, you know, pager duty or whatever to... Just because, I mean, everyone's gonna try to be more efficient, and you're supposed to see ROI productivity because if you don't, then what's the whole point of this?

Exactly. Exactly. And I, and I think it's funny, you're going back to the point of, you know, you could, you could add AI on top to solve the problems that the AI introduces, and, uh, but you just keep...

You, you, that's, like, an endless chain. Uh, and so-

But no, I mean the, the, the CodeRabbits of the world, the, the Graphites of the world would say: "Yes, actually you can."

Mm-hmm.

And so that's the little bit of the, the tension there.

Yeah. You know, I, I, uh, I've been thinking a lot about how the craft of, of engineering is evolving and, and I will say that I feel further away from being able to predict what, what it looks like than I, I did this past summer when I spent a bunch of time, um...

I actually basically went on leave for a month and joined the, um, joined the, the, the team that, uh, the, the AI team that we were building just to go and build alongside them. And I felt like it was really important for me to deeply understand, uh, the problems and the tech.

Uh, but, and so that was me. I was, I was, you know, writing, pushing code, um, effectively nine, nine, six. And, uh, and I, I went through so many different moments of realization of like: "Oh, my God, this is going to change everything," to, "Oh, my God, this is just amplifying all the good and the bad in the industry," uh, to, "Oh, my God, engineers are not gonna have a job anymore," to, you know.

It's like- And so it, it, I, I don't have any pred-- like, I felt like I had all the predictions back then, and at this point now, I'm just very interested to watch the, the phenomenon continue to unfold in front of us.

And, uh, I will say I was chatting with a bunch of really bright, uh, you know, college juniors and seniors at a dinner we hosted last night. And, um, all these folks are about to enter the industry basically having kinda come up in the, the era of agentic development and LLMs.

And I asked them like: "So what is your workflow when you're, like, building, uh, like, building a project? Uh, how do you, how do you use agents versus, like, when do you decide you're gonna actually just write code by hand?"

And I was surprised to hear the consensus was that most people there were using agents to collaborate on, like, building a design document and, like, collaborating on the architecture of the solution that they want to build, and then maybe asking it to, like, init, uh, you know, a doc or an implementation plan, but then they'll go and write a lot of the code themselves still.

Uh, so it's a little bit more of the, the, uh, the rubber duck co-architect, uh, uh, use case that was most prevalent in that group. I, I was very surprised by that.

I, I'm impressed. The kid- the kids are all right.

Yeah, I know. No, they still wanna, they still wanna actually write the code themselves. It's interesting.

Yeah, what we hear from, like, the Gen Zs at OpenAI, they, they just YOLO everything into Codex and-

Um, yeah, I would say most of the code I generate is like, yeah. But, but I spend a lot of time on the doc. It's curious, like, when you're, like, younger in your careers, like, you, you don't really have all the mental models of the different patterns to instruct.

I feel like there's, like, over-reliance, especially if you're doing the design doc, you know. I, I feel like most of the senior engineers will spend more time on that. It's like even things like, you know- What column should you index-

Mm-hmm

... depending on, you know, what queries we usually run on this table and things like that. It's hard for any AI to know that.

Right.

You know? And it's like, I feel like the, the role of, like, the more senior engineer should actually be more of this. It's like spending time b- teaching the AI, and then the AI can teach the junior people in a way.

Yeah, yeah. And it, it, everything, everything looks like mentorship and management-

Right

... at the end of the day, right? It's like you're breaking down tasks, you're, uh, you're supervising work, you're giving feedback. Like, it's, you know, it's basically management.

Except that there's agents are really bad at memory still. Like, they basically have zero memory. And then it's, it's, it's the end of 2025. What's going on?

Yeah.

Yeah. What's your internal stack for like, uh, uh, preferences? There's like kinda like, you know, explicit preference you can use with, uh, you know, agents.md and all that stuff. Uh, there's implicit preference with linter rules and things like that in a way where it's like it just happens, you don't have to tell it.

Uh, how do you structure that?

Oh, now you're talking about for agentic coding or-

Yeah

... memory within-

Just, just like-

... like a platform?

Yeah, yeah, for like the coding specifically. It's like... And then we can kinda talk about, you know, the whole Brex platform.

Yeah, just, just, it, nothing, nothing special. Just a lot of, um, like explicit rules.

That MD file.

Yeah. And then we have, uh, and we, um, in linting, uh, uh, we still have like traditional linters in place for the couple of different language tool chains, and then we're, we're, we're big fans of Greptile, and we use them for basically all of sort of the, um, smarter than linting, uh, like agentic code review.

Uh, that's been the one solution that we've aligned around that has served us extremely well.

Yeah.

Could I-

Go Greptile.

Yeah, no, we're, we're huge fans. They're-- They've built something really impressive, and I think the thing that constantly blows my mind about it is, um, the way that they're able to just have a really impressive signal-to-to-noise ratio. Like the, the comments that it leaves are very, very high signal.

Uh, like never, I never regret going through all like 65 comments it leaves on my, on my diffs because it, it catches so many things.

Yeah. I found the Codex review to be really good. I don't use Codex for code generation, but like the review product-

Yeah

... is like very good for some reason. Um, I used to have, when I was working in Rails, there was like this project called Danger Systems.

Oh, yeah.

It was kinda like-

Yep

... a semantic linter.

Exactly.

I feel like there should be more of that now. It's kinda like the rules are one thing in generation, but I want something in my CI that is like enforce these rules and call out where they're broken.

Mm-hmm.

And then I can just copy-paste that, uh, in an agent, but...

Yeah. When we, when we started building this, this new agent, um, code base, like 'cause w- as I was saying, like we were answering the question, what would you do if you built a, you know, a Brex disruptor today?

And it's like, it wouldn't be to pick Kotlin and Elixir-

Right

... as the back end. And, uh, and so we actually went with the full like TypeScript stack, and, and we, we were building on all like public interfaces and, um, really trying to make sure that this agent layer was, uh, like arm's length from, from the, the good and the bad of, of the core of our product.

And, um, and one thing, I think what we did early on, and I don't actually know if this is true, because again, the team keeps-

Right

... sorta iterating. Uh, but we were having good u- uh, good luck using, um, Claude Code, like in a GitHub action to basically go and do, uh, do more of that Danger-style like code review. So have a, a, a prompt for it that went through all of the different facets that were more conceptual versus like rigidly enforceable by a linter-

Yeah

... uh, and have it leave a big comment at the end with, um, you know, your conformance to the idiomatic coding patterns of the, of the new repo.

I wanted to spend some time, you said you wanted to deep dive on operational agents.

Mm-hmm.

Uh, customer support, onboarding, KYC, fraud, delinquent account disputes. Uh, this is how I imagine the bulk of it-

Yes

... of, of the work. Anywhere where there's a good story about maybe, um, when you started out, it was, it was gonna be this way, and then you discovered through building or through customer contacts that it had to go a different direction.

And so that difference in beliefs is something that people can learn from.

The, the thing that immediately comes to mind is that we, uh, we believed at the beginning that using RL for credit decisions would actually be a, like would be the way that we would end up... Or like credit and underwriting, like how much of a, of a limit should we give to this business, um, that reinforcement learning would be the way that we would go about, um, building a model that effectively would decision in the way that, um, a human underwriter would.

And it turns out that it was, we made this big investment. We were working with some outside, uh, like the, like a company, uh, that specializes in this, and the performance we ended up getting was inferior to just building a, like a web research agent.

Yeah.

And so, so I think what, what we took away... Uh, what, what has been most evident in operational AI is that in operations, you need to be able to break down problems really granularly and be able to form SOPs that humans can repeatably follow and, and thus, and be audited.

Uh, because so much of the, the responsibilities in operations is to, uh, is to have audible, repeatable processes that help to ensure that we're operating in a compliant manner. And that actually translates just so cleanly to LLMs that we haven't needed to use too many sophisticated techniques in, in operational AI.

Uh, it's been a, it's been relatively simple, like new tool, uh, like agents or maybe even a lot of problems can be solved with just like a single turn, uh, chat completion. And so the fact that we didn't...

Well, we did one, one sort of attempt to over-engineer and use more sophisticated techniques, uh, and we, we, we discovered, uh, that in fact the solutions are a bit more, uh, more plain and, and less technically sophisticated. The, the challenge is really articulating and refining prompts to reflect, reflect the execution of the SOP and like reflect all of the sort of institutional knowledge that isn't written down, uh, so that, uh, agents can properly replace like the, the humans or the contractors we would have making these decisions.

How do you decide what is worth like spending a lot of time building versus what you think some of these models are just... Because some of these tasks are so generic-

Mm-hmm

... they're not really about Brex.

Yep.

Like you can assume the models will be good at it, versus some of them are like very specific to you.

We kind of prioritize like the, the tasks that are most common for the broadest number of customers, and the, uh, some of them are, are, are fairly, um, fairly intuitive, like being able to research, uh, a customer to look to assess like legitimacy of the business-

Mm

... and whether that business would fit our ideal customer profile for, for onboarding, because there are certain types of businesses that we either legally cannot serve or we are not comfortable being able to serve. So that's the type of really kind of basic research, um, and like a, a relatively straightforward problem, uh, that isn't hyper Brex specific.

The things that are a little bit more specific to, to us or, or companies in our sector would be preparing documentation for a network card dispute. Like if, if you go and dispute a transaction on, on your, your personal card, you will provide evidence to your card issuer.

The card issuer then has to put together like a three or four-page Word document that goes to, uh, the card network and then eventually goes to the acquiring bank. And, and all of that is like much more specific to our business.

It's a huge operational overhead for us, and that's something that we, we decided to automate later because it's not as, uh, it's not on the critical path of like serving the vast number of our customers.

Right.

Like disputes are expensive, but not very common operational process, and so they're lower on the stack. And, uh, I think we're, we're getting there right now. But this year has basically been us just kind of like looking at every single process, just kind of stack ranking.

And, um, I will say like the thing that got us started down this path, um, was we wanted to expand our ideal customer profile to support more busin- like a wider variety of commercial businesses, which tend to be businesses that aren't growing as quickly.

Uh, so they're not like tech startups, which have a lot of growth, and they're not usually like, they're not enterprises, which also tend to have a lot of growth. It's more like a lawyer's, a, a law firm or a dentist office.

Uh, these types of like solid businesses that we should be able to serve and underwrite, but the cost to, to onboard them and the cost to serve if you have all, uh, all the humans in the loop make them ROI negative.

And so that was-

Yeah

... the first sort of use case of, of AI within our ops, uh, ops organization that then led to us really understanding we could automate much more than that.

Is this Brex going back into SMBs?

Ah, that's a good question. Yeah, yeah. So never, never let, let, let that die. Uh, you know, no, we, um, I think the way we've thought about this is we want to always like offer our product to customers where we believe we have a, like a, an offering that is well suited to the, to the needs of those businesses.

And I, I would say that still for very small, uh, businesses, our offering isn't, it's not built for that. It's built for, it's built for companies that have a s- some degree of scale, typically have at least sort of one person, if not a couple of people in the, their finance team.

And so we consider these to be more like the, the commercial segment. And so it rhymes with, with SB, but our approach back then was, uh, was a little bit more naive. And I would say we also, we were just going for a volumes, like a volume game there.

Uh, our, our internal controls were not as strong. We didn't have as much experience like underwriting those businesses. And so it, it was really ended up being a huge, uh, burden for the business, uh, almost existential, uh, for us to have those tens of thousands of customers that all were, uh, ROI negative.

Uh, and so we're, we're trying to basically scale to serve more businesses outside of tech and outside of like the op market segment, but, um, but do it thoughtfully. So I think right now our, um, our minimum threshold is, is, uh, like $1 million a year in recurrent, uh, in, in annual revenue or, um, or like $10,000 or more per month in, in card transactions as kind of being like the low end of our ICP, which is obviously not what you would think when you think of a small business.

Like small businesses tend to still be smaller than that.

Oh, wow. That's really small. Okay.

Yeah.

Yeah. Mid-market.

Yeah, exactly. And it's funny, it's just like the, the, the names of these segments, um, you know, it's like you, what, what-

Is that we call it? I don't know.

Yeah, no, I think, I think like that's, it's like, yeah, it's like lower mid-market. And it, it's funny though because when what we call enterprise may be, uh, another, you know, what sales... What we call enterprise is a business that Salesforce might call a mid-market.

Right.

Like, because it's just a, depends on the scale of yourself as a business when you use these terms.

And all of these things are built in the Brex Agent platform, like all of these-

Mm-hmm

... automations that people build?

Yes, exactly. Yeah, and in fact, the, um, most of the operational AI is running on that original platform that we have. And we, we built it, one element of it that I didn't mention is that it, um, it also, most of the UI/UX for this platform is, uh, built in Retool.

And so like you, you can basically go into Retool and, uh, there's like a, a prompt manager, a tool manager, an eval manager, um, and that's sort of where, uh, much of this was built. And the goal with that was, again, to make it more accessible, more ergonomic to, to get started.

But what an, a secondary effect of having a more like visual set of tools for this is it's enabled members of the ops organization to go and do prompt refinement themselves, so you don't need engineers to go and, and refine the prompts or, um, or even like, uh, test new foundational models when they come out.

I think that that's another, uh, fun thing when like a new, uh, when a new model drops, uh, folks will go into the, the platform and basically run the evals on the new, um, on the new model and kind of see like, can we get better performance here or does this have different, uh, different latency or different, uh, like cost characteristics.

Yeah, you want the domain experts or the people directly using the tool, not the engineers who are-

Yep

... sort of somewhat removed from the tool.

Yep.

Uh, yeah, I, I, I, I do wanna highlight to listeners that, uh, a lot of the Brex Agent platform are just things that every company should have.

Basically, uh, problem management system, which we talked about, where the domain experts are doing it. Multi-model testing, evaluation, and benchmarking frameworks, API integrations for automated workflows, NCP-based archite-architecture, shared with Brex's external AI products. This one is obviously very Brex specific.

Uh, one thing I did want to highlight that I was semi-impressed by because nobody-- people very few rarely talk about this, is knowledge base for understanding Brex's business.

Mm. Yeah.

So do you wanna expand on that?

Yeah. And th-this is an area where we've only scratched the surface here. But-

Yeah

... but a big, a big challenge that, that we face is that the world knowledge or the knowledge that's built into the model about, uh, about, you know, what GPT-5 thinks Brex does and how it thinks our business operates is actually quite different from what our business offers today or how our product works.

And so we've had to, to work on building a corpus of sort of product documentation, process documentation, and like curate this set of information to basically ground a variety of our LLM applications, including like that Brex assistant, which is like the, you know, the assistant that employees, uh, will, will talk to.

It's like we don't want it to, to hallucinate features that we don't have or like give, give wrong information there. And similarly, like, uh, some of the operational, um, uh, agents need to be grounded on, um, like what our ICP is.

Because if you ask, uh, you know, ChatGPT-5 right now, like what types of businesses does Brex, uh, onboard or like what types of businesses does Brex serve, it might not give an accurate, uh, explanation to that, to that question.

It might, it might say, "Hey, we're a corporate card for startups," which is what we did, you know, seven years ago. And it might say, "We're only-- We only serve enterprises." And so that has been an interesting challenge, and I think we're-- what we've been trying to do there is I'm actually going to be spending time with, with folks, uh, talking about this next week internally about like, can we refresh our strategy and kinda unify it?

Because we have a lot of product documentation that's internal for like our operations and go-to-market teams. We have a bunch of product documentation that's external for our customers. We have a lot of, uh, go-to-market, uh, sorta enablement material that's more, uh, sales pitchy.

And, um, we have documentation that is put into Sierra, which is the, you know, the, the chat assistant that we use for, um, uh, for frontline, um, support. Like all of this ideally could draw from the same source, but right now it's, uh, right now it's a little bit fragmented.

It's just something that we're trying to invest in though, because I think at the end of the day, the duplication, uh, of, of efforts is just like is, is, is wasteful and it's absolutely necessary to, to get this right.

Uh, just to deduplicate-

Mm-hmm

... uh, Sierra meaning the Brett Taylor startup.

Yes, exactly.

Okay.

Yep, so Sierra-

I would expect that you have-- you build so many other agents, that's, that's one you can build yourself.

That's like solving problems that are not differentiated enough, uh-

Okay

... for us. I think what, what's interesting about the, the Sierra that has been really helpful is that again, it's really easy for like the UI and UX of basically administering a Siera-Sierra agent is something that's really accessible for the ops, uh, and CX strategy team, which are like it's much more low code and, uh, more sort of workflow and DAG oriented.

And the-- we have engineers kinda going and giving it tools to, to take actions. But for the most part, like it's nice to not have to build, build the UX for somebody to manage something like that. And I think the fact that Sierra speaks the language of, of-

Few

... customers-- Yeah, exactly. Speaks the language of CX. They can do all the reporting and the telemetry and stuff that, that, um, our, you know, VP of CX, uh, would like to see. Just, you know, it's just one fewer thing that we have to build.

What about, um, evals? How do you build evals? Who manages them?

Well, it depends on, uh, it depends on the application. So on the, on the operational AI side, um, those evals are, are basically baked into the, in the platform around every, um, every prompt or every agent. And for the most part, I think most of these use cases kinda come online, like the V1 of like our, our, um, commercial underwriting agent or the V1 of our, our startup KYC agent are co-developed between like a subject matter expert in ops and like an engineer, and they're gonna kinda co-develop, um, uh, an initial eval set.

But then from there, generally in ops, you're always doing QA, be it like on humans or on, uh, on, on the LLM, uh, decisions. And so whenever, like as part of our QA feedback loop, whenever there is, uh, a mistake, that's usually almost always gonna result in like, uh, another eval being written as like a regression test.

Uh, so all of that within ops AI is pretty, pretty straightforwardly managed. On the product AI side, that's where it starts getting a little bit more challenging because the multi, multi-agent network, um, is quite challenging to evaluate. And so what we do there is we try to adopt some of the state-of-the-art for multi-turn evals, where we will, um, will basically have a, an agent embody the user and like, you know, have basically the, the end user agent is given an objective, and then we basically have it run a multi-turn, um, uh, conversation and then use LLM as judge at the end to do all of the different, uh, asset assessment.

The one other thing that we do technique-wise that is interesting is sometimes you want to-- you don't want to do like, you know, I think these multi-turn evals are kinda like integration tests. They, um, they sometimes test more than, than you-- what you want to, to assess.

And so sometimes what we'll do is we'll also pre-can like an initial preamble to a conversation or maybe a couple turns will be handwritten. Then we'll basically set, set the, the, um, eval to start, and we'll see if, uh, we're able to like isolate certain, um, certain behaviors.

So, uh, it's, it's still like a work in progress. And I say like at the end of the day, a lot of the just, um, periodic rev-- human review and, and like looking at, um, at cases where, uh, we've detected as we go to like summarize, um, like what we'll do is we'll reflect on a conversation after a certain amount of time has, has passed where we'll summarize it, like extract Assets like did it seem like the user accomplished their, their objective, and we'll just manually when, uh, a lot of the cases when that's, that's failed and sidestep an eval for it.

Mm-hmm.

Are all the evals supposed to pass or do you have a set of evals that are like, someday the model will be good enough and like-

Oh, yeah

... how could that change over time?

Yeah, it's interesting. Um, I don't know if we have any that, that are like, "Oh, someday I hope it'll be good enough to do this," but it's like there, there are the evals that are, are blocking because they would indicate like a, a regression, an unacceptable regression.

So these tend to be just accuracy related, um, evals, but then there are others that are more about like tone and, uh, coherency and these types of things where they're, they're more subjective, and we were just looking at those over time as a-

Right

... as a metric. Uh, but the, the team is actually interested, and I think we're gonna get an, a big update on like how the team is thinking about evals tomorrow and like our Friday, uh, our Friday review.

So it's... this is an area where I'd say the largest challenge, like the largest change we needed to make in how we were executing sort of as like a lab or an incubator back, uh, earlier this year to like where we are now, where we've, we've shipped and like we're trying to, to increase the rigor has been around, uh, like avoiding regressions and having more and more increasingly robust evals.

Yeah, I've worked with a company called Verase AI that does user simulations-

Mm-hmm

... and I, I, I think like that's what's been interesting. Some of these things they just don't expect, like the customer does not expect the model to do-

Mm-hmm

... but they wanna track-

Yeah

... the saturation of the model in a way, if that makes sense. And I feel like most companies know what they don't want to happen.

Yep.

But it's almost like they don't... they cannot quite articulate, "Oh, I want in the future the model to be able to do this." They can do it today, but I'll, I'll keep running this eval.

That's actually really, really interesting to me, and I, I, I'm gonna take that away and, and, and start thinking about this because there are, there are going to be s- certain... I mean, we already seen this where, where, uh, users will ask the assistant for help with things that we don't support yet-

Right

... or we haven't implemented yet. It's like those are opportunities actually for us to build a, like e- effectively write a test that's going to be fail-- like failing-

Right

... for, for weeks or months and, and eventually will go green. But as a way for us to actually kind of show like the progression of sophistication of the assistant. I, I really, I really like that as an idea.

Yeah. I, I wonder how you also catch hallucinations and-

Yeah. Yeah

... of things that it doesn't have.

That's usually the, that's usually the problem is it-

Yeah

... it'll, it'll, it'll, it'll pretend like it can assist with something, and it'll, uh... Like, one thing that is really annoying that has been tough to, um, to prevent is that the, the assistant, because it is used to speaking to other agents, um, that can support it in like accomplishing various tasks, if you ask it to, to help with a task that it thinks it probably should have an agent to, uh, uh-

Mm-hmm

... to, to work with, it'll just hallucinate that it, you know, it's like, "Oh, yes, I'll, I will like, you know, I'll reach out to the finance team on your behalf to, uh, to pass this question along," but it's not doing anything.

There's like no finance team. There's no way for it to do that. This is something that comes up a lot. It's like, "Would you like me to ask the finance team?" And, uh, there's no, there's no actual tool for that.

Do you put guardrails for that?

Yeah, yeah, that was, that was something that we had to, um-

Like a regex, like-

Oh, no, we don't... I think we've been a- we've been able to just beat that out of its system with a system prompt. But, uh, but the... We don't have as many guardrails in place right now, just around a couple of like potential, uh, like things that could get us into trouble-

Yeah, really extreme ones

... with hallucinations. Yeah.

Yeah. I, I just... Yeah. It's surprising when I, I guess two years ago was first kicking around the idea of all these things, I would have s- said that probably guardrails would be more prevalent-

Mm-hmm

... especially in finance use cases, but surprisingly, they're not.

Yeah. And that was actually part of what we... That was like a feature I believe we built in the LLM, LLM gateway early on is like the, the sort of last chance like, uh, um-

Like hard-coded

... guardrails. Yeah, exactly. Here's some regexes-

Everything is-

... that will just kill... Yeah, exactly. Or just, you know, in the way that like if you go way afield on, uh, ChatGPT, you just get like the inline 500 error. It doesn't even tell you that it can't help.

It just like craps out. Uh, like I... we kind of built a couple of those circuit breaker, or like the ability to put those circuit breakers in and, and I don't, I don't believe we're using them for anything.

One last thing I want to get your thoughts on was AI fluency levels-

Mm-hmm

... which you guys have a framework of user, advocate, builder, native, and everyone goes through it, including Camilla.

Mm-hmm.

And I just think it's interesting. I think it's a model that other people are thinking about adopting, but they're worried about rolling it out. Um-

That everybody's gonna be bad and then I let-

Well, and also like how do you have like this in-house training course that you keep up to date? Yeah, just t- tell us more about it.

Yeah. So in, in the operations org, uh, they're actually more ahead of even engineering on this front as far as like trying to create, um, create like learning pathways for this. Uh, and I think that part of the reason why they're ahead of us is that in operations, they're much more, uh...

they have to be able to operate, uh, training at, at scale. Like training is a big, very big part of, um, of how people build aptitude around their, their job function within ops. Whereas like in EPD, a lot of it is sort of, uh, getting hands-on, building experience, like going a lot and getting mentored, getting code review.

Mm-hmm.

But, uh, it's been really neat because I think we've really like... We created an environment. We managed to, by, by speaking openly a- about the, the transformation that we saw would happen in this industry towards AI sort of displacing a lot of, um, a lot of the operations and CX roles, and we were just honest about it.

And I think what, what in the same breath that we said, "Hey, a lot of these job responsibilities will go away," we also said, "We don't anticipate that meaning that your job has to go away. It's just that your job has to change."

And so the, the fluency framework and then the, the like the training and support and like the positive sort of culture where we celebrate people making progress i- has been really helpful for like avoiding a culture of fear or like, "Oh, you have to do this," or, "You have...

You're gonna get... This is gonna go in your performance evaluation." I think the-

It does

... it... Well, it's not like it's rote. It's like, oh, like what is, you know, what is the... like how much are you using AI, and is it, is it enough? It's, it's more- I think we've built a pretty, like, positively framed culture where we'll, we'll do like spot bonuses for, for people who have like particularly novel uses, uh, of AI on-- in their day to day.

Um, in our company all-hands every two weeks, we'll do an AI spotlight, and it's very rarely somebody in EPD. For the most part, it's folks in ETMs, ops, uh, finance, the people organization showing off like how they're building agents, you know, in ChatGPT or on Glean or how they're-- they like just found some new use case that they thought was helpful.

So we're trying to create, create like, um... I think at the end of the day, like we've hired a bunch of really smart people who I-- like, I have full confidence that, that, that this type of work is in within the reach of anybody who's motivated to like sorta challenge themselves.

And so we've, we've done that. Then in engineering, there's one other thing that I wanna call out 'cause I think that this is kind of fun, is that we adapted our e-interview loop to be more AI sorta agentic coding native.

So instead of, um, we had, uh, like a coding and a system design question that we basically have revamped into, um, a project where we'll give you like a brief before you come on site and then like an additional sorta spec when you do, um, when you start.

You know, we expect you to use agentic coding to complete the, the task. In fact, it's like kind of impossible to get all the way through it if you don't. And so we're evaluating, you know, your knowledge. Like we're kinda watching how you work.

We're evaluating whether you understand the code that's coming out. We, we, you know, we're kinda probing at you as you go. But what we did in order to kind of maybe bootstrap the process of all of our existing engineers like getting familiar with agentic coding is that we, as soon as we had the interview, um, ready to ship, we started...

We said everybody in engineering, including all the managers, are gonna have to go through this interview. And so we re-interviewed everybody internally, and it's like, it's one of those things where it's like it, it's not a... We didn't like keep a score or like, or like, you know, I don't have any data on like who passed or failed or what they- ...

what they scored. But what we found is like, as people would take it, it would actually cause them to have moments of realization where it was like, "Oh, I, I can uplevel my skills around this," or like, "I have-- like I want to be better at this."

And so, uh, we're trying to find like, um, a way, like a variety of techniques to kinda push the culture along. Uh, and I think as I reflect on like the year, 'cause this is the year where we really put all the effort into it, um, I'm really satisfied to see the extent to which everybody's leaning in on a, on a daily basis.

Going back to like even I, I was shocked when we were looking at our cursor logs that like the number one user is the, is an engineering manager on Infra, Infra org. It's like that, that is super cool to me.

It means that like folks have, have, uh, have taken this to heart-

Mm-hmm

... and found, found ways of, um, doing their job differently.

I guess my... I, I had a closing question or I guess a parting question, a-and this is broadening out from Brex.

Yeah.

And this is just you interface with other engineering leaders all the time.

Mm-hmm.

Did we not cover anything that other CTOs are having as top of mind today? Like their number one problem is underscore.

The thing I find myself discussing with, with folks that... And I, I don't wanna shy away from like scary topics. Uh, in fact, we're just, just kind of on one that was adjacent, which is like how do you evaluate somebody's like progression towards being more AI native.

Yeah.

The, the, the, the cousin to that question is it's like will we need as many people, uh-

Yeah

... to operate our businesses? Like are there layoffs coming? Are-- how are we, how are we thinking about like, um, headcount growth? Uh and-

Junior versus senior

... junior versus senior, yes, exactly, like level mix. Um, and I still have more questions than I have answers there. I think, I think what has been really interesting is that I view agentic development as being something that amplifies all the, all the good just as much as it amplifies all the bad.

And it amplifies, uh, uh, sloppiness, poor architectural thinking, um, uh, misunderstanding of, of the requirements. Like there are, are... for all of the, the acceleration of good outcomes, it also accelerates bad outcomes. And like what has been interesting is that there has been, when you sum that all together, there's less of a obvious, um, like capacity increase.

It's, it's more, it's more nuanced than that. And so I'm not looking at headcount planning, uh, as, as we think about it next year as, as being something like-

Yeah

... oh, well, because AI is giving us so much more leverage, we, we don't need as many people. Um, we've actually... The thing I'm really proud of in, in my tenure as CTO is that we, we haven't grown engineering at all.

What we've done is we've, we've grown the business significantly, but we've been able to build, uh, like greater efficiencies in, in how we execute, like how we, how we, uh, we think about building, how we roadmap, um, what we choose to do and what not to do that we're able to, uh, to serve significantly more customers with more lines of business, um, without needing to grow engineering headcount.

I think that that's kind of the way that we're gonna just continue on this road is like I like having 300 engineers. Like I would love, love to just, you know, a year from now have 300 engineers, but we're still, you know, 30, 50, 100% more efficient.

Uh, uh, that, that is the thing that comes up with, uh, with other engineering leaders. And the other part of that conversation is like how much is AI getting blamed for just sort of ordinary performance-oriented, uh, risks.

Yeah.

You know, like if, if Microsoft is letting go of like 4,000 people as a business, what they have 150,000 employees, I believe. Uh, is that really like AI causing that or is it them just using it as a way to, uh, to avoid some harder like perf management decisions?

I'm not entirely sure, but I'm, I'm listening more than I'm speaking on the, on this topic because, uh, every time I feel like I have a pretty firm point of view, some new, uh, anecdote or experience comes in that kinda challenges or invalidates it.

Yeah. Well, you know, it... I, I take these signals as it's my job to go find people who think they have answers and surface them. And you may or may not disagree, but at least you have something to use as a straw man in, in your work.

Exactly. Exactly. And I, and I think as, as an industry, we're just, just early innings on, on, on this transformation. So I'm looking forward to seeing, uh- Uh, you know, listening to this, this podcast episode a year from now and, and, and seeing, you know, what we got right, what we got wrong, and what's different, uh-

Yeah

... because so much changes, uh, quarter over quarter.

Yeah. I do think AI COE is a very well-established pattern. I think, uh, internal platform is a very well-established pattern, and this, uh, fluency thing is something that people are figuring out that I think you guys are a hit on.

I'm happy to hear that.

That'll be my feedback.

Yeah.

Any final call to action for things that you wanna buy? Like what should people build for you, like problems you're trying to solve that you would love people to reach out for to, to help with?

The call that I'd make is for folks who are interested in, in multi-agent networks to, to get in touch with us because I, I do feel like this is something where, where we're, we're innovating in, in service of, of our customers and where I, I feel like the, the frameworks, the tooling, um, and the, the research is, is, is there.

There's actually quite a lot of like interesting papers and things that w- we lean on. Uh, but I would love to, uh, would love to see more of that like en- encoded in the, um, in the... what's available writ large in the industry because I feel like my intuition has been that trying to graft LLMs into deterministic workflows and DAGs is, is kind of underselling like the power that they have to actually plan and execute m- more, in a more sophisticated, like fluid way.

And, and I, and I just want to see like the, the industry lean in more, um, on uh, on these agent-to-agent, uh, uh, interactions.

Okay. So I, I, I'll dive in a little bit here-

Yeah

... 'cause I, I have a minor opinion. You keep using the word networks.

Yep.

Is that a reference to a specific paper, or it's your term for it?

It's just our, it, it's our term, and I think that that is, that's actually the term that Master uses as well, um, for it. It, it, we, um, yeah, initially we used to call them agent run times, uh, internally, and then we just, yeah, switched to networks.

Uh, and then I think the other thing I wanted to get a clarification on is, is it mostly a full agent talking with a full agent, or is there a kind of like a orchestrated boss agent talking to a sub-agent?

And I think that does matter for a subset of people who are building all these things because when you say multi-agent-

Yeah

... sometimes people don't agree what that means.

Yeah. So it's, it's a tree more than it is a graph. So it is like, yeah, we have-

Just 'cause when you say network, it's feels more of a graph.

Yeah.

But it seems more directional as a tree.

Mm.

Like there, there is a hierarchy.

There is a hierarchy, yeah. But there, but there are some violations of that. Like one of the, one of the interesting use cases, uh, and this is where like the power of, of having an, an assistant for every employee, plus having agents that run and, and embody, uh, members of the finance team is really powerful because a, a, there's this interesting use case that, that we brought to market, which is that, um, one of the finance team agents that we, we, uh, launched is an audit agent, where like an audit agent kind of embodies the work that a lot of larger finance teams will do to look for patterns of waste, fraud, or abuse or like systematic, uh, avoidance of policy that isn't as obvious with a single expense.

Like you can evaluate a single expense and the metadata around it to see if it, if it's, um, within policy or not. But, um, what if you start seeing an employee often make a, a large number of like $74 transactions when receipts are required at seventy-five?

Or what if you, what if you see, um, certain things like, okay, there's actually a fair number of like DoorDash expenses during business hours from this individual, like on, on days that an office lunch is provided, or maybe you see like rideshare patterns that are, are, um, w- where you have to look at a broader context.

Um, so we built this audit agent that can like ingest your SOP and, and look also ingest your-

This is a Brex's customer's SOP.

Exactly, yep. And, uh, and what it does then is it's, it's basically always looking for potential violations, and what it does is it, it is extremely zealous. Like it, it wants to have a minimum number of false negatives, so it will raise a large number of potential violations.

And then a separate agent, a review agent, will then apply wisdom, the wisdom of like, is this important enough to follow up on? Is the dollar amount in question high enough? Does this user seem to have like a high compliance behavior more generally?

It makes a judgment call about whether it's worthy enough to take that violation and make it into a case. Then once it's made into a case, generally what happens is that you need to get more information from the individual.

So if humans were doing this, there'd, there'd be some outsourced team that's like looking for all the potential violations. Then you have some full-time employee on the finance team who's, who's looking at all the violations, "Oh, these are the ones that are important.

We need to follow up on it." Now what they do is they hand it off to somebody who will go and Slack that employee and be like, "Hey, what's going on here?" And so what we have is like the audit agent looks for violations, the review agent decides whether it's worthy enough to turn into a case, and then from there, uh, when the case is filed, the, that, that will trigger an event to the Brex assistant for that employee.

And like any additional information about like the business justification, um, can be collected. Or maybe the assistant already knows because it, in its conversation history with the employee knew something about why this, this expense, uh, looked out of, out of policy.

And so you start having... Then the network becomes interesting when you have the finance team agents communicating with, uh, the assistant for various employees, and then behind there you have other, other sub-agents, and so then you start seeing like more of a graph, uh, emerge.

But when you look at just what serves the employee, it looks more like a tree.

Amazing. Well, I didn't know you were gonna go into that level of detail.

Yeah, yeah, sorry about that.

No, no, no. I'm, I'm actually really glad I asked. Like that is very impressive and, uh, I hope you, uh, do more content about that.

Yeah, absolutely. We're, we're really excited about it. I think, uh, it's, it's been, it's been good to finally figure out, uh, a use for, for agents and have the technology be as, uh, like as sort of robust as it is to start realizing this vision, 'cause it's something that we, we kinda dreamt of a couple years ago.

And the tech, like to your earlier point, the tech just wasn't there when we were trying to make the, make the, a similar concept work with GPT-3.5. It was like, "Nope."

Right.

We were hallucinating tool calls in, uh, back in that day.

Um, awesome, man. Thanks so much for joining us. This was fun.

Yeah. I really enjoyed it. Uh, happy holidays, guys. Thank you for having me.

Thank you.