How Braintrust uses AI agents, evals, and CI to ship better software | Ankur Goyal

Summary

How Braintrust Uses AI Agents, Evals, and CI to Ship Better Software

How I AI with Claire Vo | Guest: Ankur Goyal (CEO, Braintrust)
Date: June 15, 2026

---

Episode Overview

In this technical deep dive, Claire Vo sits down with Ankur Goyal, CEO of Braintrust, to discuss how advanced AI agents, rigorous evaluation frameworks (“evals”), and robust continuous integration (CI) pipelines are revolutionizing software development. Geared towards senior engineering leaders and practitioners, this episode showcases not only how AI can tackle complex infrastructure and architecture work, but also how to operationalize benchmarks and feedback loops that deliver significant quality and speed improvements.

The conversation spans from practical agent workflows and eliminating backlogs to demystifying evals and building a culture of rigor—sprinkled with live product demos, personal anecdotes, and actionable advice.

---

Key Discussion Points & Insights

1. The "Agent Line" and Reimagining Engineering Work

• Agent Line: Tasks that once required human intervention ("above the line") can increasingly be delegated to AI agents ("below the line") as models improve. The agent line “keeps going up” (14:15).
  • Ankur: "Everyone should take a hard look in the mirror and reevaluate how they spend their time. There's a lot of interactions ... that fit below the agent line." (14:15)
  • Claire: Emphasizes making time for true “maker” work by leveraging agents for repetitive or computationally intensive tasks. (13:47)

2. AI for Technical Problem-Solving: The Coding Agent Workflow

• Use Cases: Ankur details how Braintrust uses agents to routinely optimize database performance by exploring combinations of indexes, formats, and execution engines.
  • They systematically replicate slow production queries, then have agents try exhaustive solutions from database literature—far more than any human could benchmark alone. (03:47–06:08)
  • Ankur: “We're exhaustively trying every open source column store format out there ... and sort of computing the matrix of this.” (05:40)
• Risk & Adoption: Historically, infra changes are avoided due to risk and inertia, but agents allow for rapid iteration at lower cost and risk—engineers can “bite off much more interesting technical challenges with AI as your, you know, sidecar than you could before” (12:32).
• Benchmarking: Agents enable continuous, rigorous benchmarking and finding real bottlenecks and tradeoffs, e.g., Bloom filter-based indexing discovered after a week of agent-driven experiments. (09:31)
  • Ankur: "Now there is no excuse to not do those benchmarks." (09:27)

3. Practical Agent Management & Developer Experience

• Foreground Agents: Ankur runs multiple “foreground” agents simultaneously on his machine, each in its own TMUX session, to parallelize experiments and maximize productivity. (14:15–15:50)
  • Details some common developer headaches like port collisions and how community tools (e.g., Vercel’s Portless) are improving quality of life.
  • Ankur: "Right now I have four things I'm working on. So each one is a TMUX session. They're named BrainTrust 1 through Brain Trust 4." (15:00)
• Remote Experiments: For large-scale data workloads, remote agents run high-concurrency IO tests against cloud infrastructure to simulate production conditions. (16:30)
• Making Time to Code: Both Claire and Ankur emphasize fiercely protecting “maker schedule” time and minimizing meetings to achieve flow states—often aided, not hindered, by the use of agents. (19:48–20:32)
  • Claire: "All afternoon I'm just like in my real state, which is hoodie on bad products." (19:50)

4. Evals Demystified: What They Are and Why They Matter

• Concept: Evals transform product requirements into testable, quantifiable outcomes—essentially encoding “what” good looks like, rather than “how” to achieve it.
  • Ankur: “Evals are a methodology for you to say, this is what success looks like. In my opinion, evals are actually the modern version of a PRD.” (23:47)
• Practical Example:
  • Building a question-answering agent for Braintrust’s docs: Ankur uploads real user questions and uses AI to generate both the answering prompts and the scoring functions—then executes bulk evals to quantify output quality. (26:07–30:21)
  • Ankur: “It's a pain in the ass to write these criteria out by hand. So it's really nice to just let a model do it for you.” (29:38)
• “Vibe Checks”: Human judgment—what Ankur calls a “David eval,” referencing his designer—is still crucial for qualitative review, but AI-driven evals capture and scale expert taste. (31:27–33:00)
  • Ankur: “We’re able to have David’s palette applied to more things. ... The quality bar that we’re able to hit is higher because we’re able to get more things to that bar.” (33:00)

5. Continuous Integration (CI) as an Accelerator

• Velocity Through CI: To manage increasing feature throughput and avoid regressions, Braintrust invests heavily in CI. Fast, reliable pipelines become the “platform” on which agents and engineers iterate safely.
  • Ankur: “We spend a lot more time working on CI than we used to. ... If we're constrained, we don't ship crap—we pause and improve CI so we earn the ability to move faster.” (36:06)
  • Claire: “If you want to accelerate engineering velocity with AI, fix your CI.” (36:16)

6. Healthy Team Practice and Avoiding Burnout

• Time Chunking: Don’t fall into guilt or burnout over constant agent activity—chunk time intentionally for agent monitoring, coding, and meetings to optimize productivity and wellbeing. (20:32)
  • Claire: “Chunking your time with AI ... narrows you on the more productive pieces and is also just a more enjoyable way to get stuff done.” (20:53)
• Work-Life Boundaries: Both recount overzealous phases with laptops and agents intruding on personal time, and the need to set boundaries—even with powerful new tools. (21:24)

---

Notable Quotes & Memorable Moments

• On Benchmarking with Agents:
  Ankur: “There's no staff engineer who is running as many rigorous benchmarks and trying out different algorithms and analyzing ideas manually than someone who's using an agent.” (09:27)

• On Human Attention vs. AI Endurance:
  Claire: “Humans lose context on the problem over days. You have a decaying attention span towards hard but tedious problems. … The practical quality goes down.” (11:32)

• On the Agent Line:
  Ankur: “The agent line keeps going up … the best people are pushing the agent line inside their company by being smart about what skills they're writing and what integrations they're building.” (14:15)

• On Evals as Modern PRDs:
  Ankur: “Evals are actually the modern version of a PRD … you supplement that with examples ... but you let a model figure out the how, and you are really focused on the what.” (23:47)

• On Scaling Taste with Evals:
  Ankur: “We’re able to have David’s palette applied to more things. The quality bar that we’re able to hit is higher because we’re able to get more things to that bar.” (33:00)

• On CI and Velocity:
  Ankur: “We spend a lot more time working on CI than we used to ... if we're constrained, then instead of shipping a bunch of crappy stuff, we're like, okay, let's pause and improve CI so that we earn the ability to move faster.” (36:06)

• On Resetting and Handwriting Evals:
  Ankur: “Close the session. And then I improve the evals and then I try from scratch again ... I hand wrote the eval and then by the end of Sunday, the problem was solved.” (37:56–39:03)

---

Timestamps for Important Segments

• 03:47–06:08 – How coding agents are transforming database optimization
• 09:27–11:32 – Do agents actually improve quality? Benchmarking at scale
• 14:15–17:17 – Managing multiple AI agents and raising the "agent line"
• 23:47–26:02 – Evals explained: from PRD to quantifiable AI feedback loop
• 26:07–30:21 – Live example: prompt and scoring workflow in Braintrust
• 31:27–33:00 – Incorporating expert “taste” into systemized workflows
• 36:06–36:16 – Continuous integration as the keystone for fast, safe engineering

---

Actionable Takeaways

• Use AI coding agents to tackle the “long tail” of architectural experiments or optimizations that humans would find too tedious or slow.
• Define strong evals—quantifiable tests for success—for every important engineering or AI system deliverable.
• Invest heavily in CI pipelines; they are the enabler for rapid iteration and safe deployments in an AI-accelerated workflow.
• Don’t underestimate the value of human expertise—use “vibe checks” from your best engineers or designers to iterate and encode system improvements.
• Protect time for deep work, chunk AI task management, and set healthy boundaries to maintain team wellbeing amidst higher velocity.

---

Further Resources

• Braintrust: braintrust.dev
• Contact Ankur: @ankrgyl on X/Twitter
• Show Notes/Links: See episode GitHub notes for Portless, Context 7, and referenced Intercom/Stripe workflows.

---

If you haven’t tuned in, this episode will give you a hands-on, inside look at how top engineering organizations are using AI not just for code generation, but to raise quality bars, automate rigor, and accelerate delivery cycles—without burning out their teams.

Transcript

[A]

I'm still in, as I say, the year of our cloud 2026. I still talk to engineers that say AI on our most complicated things cannot do a good job.

[B]

I so viscerally disagree with. There's no staff engineer who is running as many rigorous benchmarks and trying out different algorithms and analyzing ideas manually than someone who's using an agent. Everyone should take a hard look in the mirror and reevaluate how they spend their time. There's a lot of interactions that you have, or direction that you're giving or decisions that you're making. And I think, like, many of these things to me fit below the agent line. I think the agent line keeps going up.

[A]

Why do you think this concept is so important to understand? And how can you just demystify it for folks who are a little intimidated by it?

[B]

Now that models are so good at actually writing code, one of the best things that we can do is create really hard evals. And if you create the right tests and success criteria for a model, then it can be really creative and it can work on this stuff in the background and actually try to improve a bunch of things.

[A]

I have a lot of people saying, wow, if I go as so far as to turn my own taste or my own skill or my own expertise into a system, I'm functionally just building my own replacement.

[B]

We're able to have David's palette applied to more things. I think the quality bar that we're able to hit is higher because we're able to get more things to that bar.

[A]

Welcome back to How I AI. I'm Claire Veaux, product leader and AI obsessive, here on a mission to help you build better with these new tools. Today I have Ankur Goyal, the CEO of braintrust. And this is a technical one. So if you're a senior or staff engineer or a VP of engineering or cto, this one you're really going to want to pay attention to. And we're going to talk about how coding agents can help you bite off really technical architecture and infrastructure work in a way that no other human engineer could before. We're also going to demystify evals for folks and just show you exactly how you can use them to make your AI products better without having to touch a thing. Let's get to it. This episode is brought to you by Guru the AI layer of truth for your company's knowledge. Here's the problem. Your AI is only as good as the information you feed it. Most companies are getting confident but wrong answers from AI because their underlying knowledge is outdated, incomplete, or just plain incorrect. Bad information doesn't just slow you down, it costs you money and puts you at risk. Guru solves this by adding a verification layer between your company's knowledge in AI tools. Instead of just hoping your AI gets it right, Guru automatically scores content for accuracy, flags outdated information, and ensures your team gets trustworthy answers every time. It works with the tools you already use so you don't have to change how you work. Thousands of companies trust Guru to keep their AI accurate and compliant. Ready to stop playing Russian roulette with your company's knowledge? Visit getguru.com to learn more. Welcome to How I AI. I am excited to have you here.

[B]

I'm super excited to be here. Thanks for having me.

[A]

So I'm going to make you laugh, but I recently did an episode about the recent GPT5.5 model release, and I know you and I use Codex. And one of the funniest comments in that post was, claire, can you do an entire episode about tech debt? And we were talking before we got on the recording. You're like, how technical and how nerdy is this audience? And I'm like, bring it on. So we are going to talk a little bit about how you approach engineering and then how you use AI to do things like optimize slow queries. So let's, let's hop in. Tell me, tell me about your approach to software engineering in the age of AI.

[B]

You know, I spend a lot of time working on software for doing evals and observability, and that's kind of shaped my own perspective about software engineering. Like now, now that models are so good at actually writing code, one of the best things that we can do is create really hard evals. And I'm not talking about like AI evals. I mean, things like, why is this query so slow? And if you create the right tests and success criteria for a model, then it can be really creative and it can work on this stuff in the background and actually try to improve a bunch of things. So one of the things that I spend a lot of time on right now is making the queries that people run in our product faster. And people can just write arbitrary queries. Like, you know, they can. There's an example someone is trying to find like needle in a haystack of some specific kind of interaction someone had in their product, and they're looking at like billions and billions of traces and they want to find like the 5,000 or something that match. And this is over like a 90 day period or something. Like a lot of, a lot of data. And that's one example of a query. And like, okay, there are all these things that you can do in database literature, like different indexes you can build and different ways you can prefetch data and blah blah, blah, all this stuff. But how do you try all those things and how do you run all the experiments required to actually do something like this? So what we do and what I've personally spent a lot of time working on is trying to figure out, you know, manually is fine, but automatically is even better. Like what are the patterns of queries that people are running that are slow. And then we will reproduce those things and use a coding agent to try out a bunch of ideas from database literature. So like download a bunch of data locally and then maybe try different. In this case right now I'm trying out different column store formats. So we use an index underneath the scenes called Tantive, which has a built in column store. But it's not that great. Like the thing overall is great, but their column store is not like that great. And so what we're doing right now is like exhaustively trying every open source column store format out there and then exhaustively trying every column store execution engine out there and sort of computing the matrix of this. And it's, you know, it's like, it's amazing.

[A]

I completely agree. As somebody who has led engineering organizations for a really long time, when you're trying to make infrastructure platform core component changes in your application, because of both the cost of implementing those being very high and then the unknown unknowns being quite risky, teams are actually pretty risk averse in terms of making big platform shifted shifts or changes to their core implementation. It's like the thing that you shipped is the thing that you get stuck with certainly on the engineering side. And what I love about AI right now, and these coding agents in particular, and then Codex in particular particular is it has been the only setup, Codex plus these GPT models has been the only setup where I have been able to set up a very similar process which is the outcome I want is xyz. We need to programmatically test against pretty long tail data structures to figure out which of these potential solutions are going to get us closer to the outcome we want. In your instance it's database query speed and latency. In my instance I was doing a very, you can appreciate this very complex data migration of like stored structured and unstructured data generated by AI. So it was all like messed up to begin with and then I had to migrate it to a schema. And so it was like schema to schema migration. Millions and millions and millions and millions of rows and lots of edge cases. And doing that as a human takes forever. You know, you could script it and you can like bang some systems against it, but then your human ability to manage those cycles and say yes, that's right or no that's wrong or this gives us indication that we should go left or that gives us indication we should go right. And so I do feel like this combination of like a very precise outcome and an agent that's smart enough to bang its head against a really, really long tail of problems with a guided sense of the technical space, it does really well. And I have not heard this on the kind of like data store side. It's really interesting, but I just think, hey, engineering leaders out there. I've had, I've been so many debates about what we're using for our data score, how we optimize performance, what technologies we should bring into the stack versus not. And you can run those like very, very iterative loops on. I'm presuming you're using production like data or real representative queries to test that. Is that right?

[B]

You can actually use production data too, but for some subset of things. And with the right engineering in place, you can just run on production data.

[A]

Yeah.

[B]

And in many ways it's a lot safer than having humans test on the production data because no one's looking at it.

[A]

Yeah. And this is where I have so many staff engineers be really, really cynical about. Does AI have a place in their coding tools? I'm still in as I say, the year of our cloud 2026. We still, I still talk to engineers that say AI on our most complicated things cannot do a good job.

[B]

Oh, I, I so viscerally disagree with that same.

[A]

Tell me why you disagree.

[B]

Well, I, I mean, I think so. I've been working on databases for almost two decades. There's not many things that staff whatever, risk averse, blah blah, blah, all that stuff you could apply to than like literally building a database if you work on a database. We recently added this like fancy index thing into BrainTrust that uses Bloom filters. And by the way, we discovered that that would be a practical solution to the problem after running like a week of continuous experiments with different types of indexes. Bloom filters kind of have a bad reputation, but they worked out to be very effective in this case. So if you build something like that, usually what happens is the very best engineers will run a few benchmarks and then you'll send it to your peers and then your peers will shit all over it and rip it apart and say, you didn't benchmark this, you didn't benchmark that. And what you do is you prioritize the top few benchmarks and then you probably bullshit the rest like, oh, I, you know, I didn't benchmark this. However, if you read the code, you'll see it's not N squared, it's log N. And so this is not going to happen. And like half the time you're wrong. Now there is no excuse to not do those benchmarks. So now you love it. Like, we don't. I'm not spending my time like sitting and like typing the benchmark code. But I'm talking to people. We're looking at the code, we're looking at the thing. We're like, okay, well we index. We benchmarked how much faster it makes the queries. Did we actually do a good job of benchmarking how much slower it makes indexing? Oh shit, no, we didn't do that. And so we actually spent some time doing that and we discovered that we were doing a terrible job at indexing it efficiently. And so we spent a lot of time on that. And I could sort of, I don't agree with this. I could some, I can empathize with the argument that models aren't good at writing highly concurrent code or they're not good at writing very performance sensitive code. But the I, there's no staff engineer who is running as many rigorous benchmarks and trying out different algorithms and analyzing ideas manually than someone who's using an agent. And even that baseline is just incredible.

[A]

I agree. And I think there's this theoretical quality and then there's this practical quality, right? In a theoretical ideal world in which we don't sleep and every time we sit down at our laptop, we, we end up writing perfect code. And in a theoretical world in which those benchmarks are, all of them are run, not just the BS ones, like in that theoretical world you could theoretically, say, perhaps in some untested case you get better quality when, when humans are hands on, but the practical application is you lose context over like the humans lose context on the problem over days. You have a decaying attention span towards hard but tedious problems. And so I do think the practical quality goes down. And, and so I tell people, like, the practical quality of integrating AI into your engineering process on very hard technical problems goes up simply because of how hard you can Run at the problem and how long and consistently you can run against the problem. And then, you know, what I was going to go back to saying, which is you can bite off much more interesting technical challenges with AI as your, you know, sidecar than you could before, again, practically, because your company can support the cost of doing so. Right. If you're like, I want to sequester all my staff engineers to solve our database indexing problems for the next year and we're just like really going to go, go deep in the weeds and we're going to test these six different open source, you know, solutions to this and we'll come back in a year and we'll tell you if we figured it out. Not that we figured it out. Business is like, no, you know, you're, you're, they're a CEO, like, no, no, thank you. But if you say, hey, we're going to have this thing in the background and we're going to check on it, we're going to make expedient progress. Sure. And we can ship other stuff while we're at it, I think that's a really easy.

[B]

Yes, absolutely. Yeah. I mean, I think the, the motto that we have now is there's just no excuse to not have rigor. Like if. And there's no excuse to not have performance. If someone complains about something, if someone complains about a paper cut in the ui, you know, whatever it is, there's just no, there's no, we don't really have a backlog. Like, there's, there's no excuse to, to just not improve these things.

[A]

Yeah. And for folks looking for we don't have a backlog inspiration, we just interviewed Brian from Intercom who said their goal is like backlog zero, nothing in the backlog, so that everything can get shipped. Okay. So we're solving really technical problems. I think this is a great approach. How are you engineering with AI? Because I love that you're still, you know, you're writing code, you're spending time on this. Any tips or tricks for how you're managing your, your fleet of agents that you think are unique?

[B]

I think that everyone should take a hard look in the mirror and reevaluate how they spend their time. There's a lot of interactions that you have or direction that you're giving or decisions that you're making. And I think like, many of these things to me fit below the agent line. And to me, the agent line is like, if I or whoever would be at the meeting or whatever. Like if we equivalently took the information that we're discussing and we just gave it to an agent. Would it solve the same problem? And I think the agent line keeps going up and also I think the best people are pushing the agent line inside of their company by being smart about what skills they're writing and what integrations they're building and so on. So once you do that, you likely have a lot more time than you thought you did. I don't take any meetings after 12. This is the last meeting of the day for me. And that means that every day I am able to, in the Paul Graham framework of maker versus manager schedule. Every day I'm able to enter the level of focus that's required to be in the maker schedule. And so I personally write a lot of code and I spend a lot of time writing code and I, I've, I haven't spent as much time writing code in, in a while and I, I really love it. So that's, number one is like make the time. Um, my, my workflow is, is very simple right now. We don't have a great background agent set up yet. I think that we are exploring various things and trying to get there. But I, I have usually five or six foreground agents running on my computer. Each one is a TMUX session. Right now I have four things I'm working on. So each one is a TMUX session. They're named BrainTrust 1 through Brain Trust 4. And you know, each of these has like some UI running and it has some services running. There are problems like port collisions, like I can't isolate everything as much as I'd like to. And I think that there are a lot of solutions for trivial software that do this. There's not a lot of solutions for complicated software yet. And I'm excited. I mean, everyone I talk to is building their own thing. I just met a startup that's like two months old and they built their own internal tool for doing background agent PRs, which is, I'm, I don't judge them for it. Like I don't know what else they would do, but it's kind of crazy. And then I also have remote ones. So here's one where I'm, I'm working on trying to improve our column store performance. And this is running on not real data, but close to real data and it's running remotely and it's, you know, it's, it's, it's running like much more scale and many. And I mean if I ran this on my computer, it would probably die from just how much compute it's using. But, but I'm, I'm able to, to in this case test like what's the real latency between EC2 and S3. If I'm trying to do like 4000 concurrent reads, is it enough, is it not enough for this workload? Can I interleave things, whatever properly? And I've been running this experiment for, for several days just trying to figure out like what's the best, you know, right now I'm, I'm talking to it about what the indexing lifecycle should be because I think we figured out how to make the queries fast enough.

[A]

Some people are going to be listening to this being like, oh my gosh, this is so technical. I don't have these problems. Let me take a step back for folks and tell you what I think I'm seeing here, which is one, you're using Codex, right? Yeah, Codex for hard problems people. I'm telling you just that I think

[B]

it's currently the only model that will disagree with you regularly. And I think if you're working on hard problems, it's very important.

[A]

And then for you, what I'm also hearing is you're using foreground agents. You can base. You basically have a personal concurrency limit of like, let's call it four.

[B]

Sure.

[A]

Which is about what, about what I can do as well. So I think people ask me all the time, how do you handle all this context? I'm like, I don't do more than I think I can do at any one time. And I also, I have more trivial problems than you. So I think you're right in that the current sort of commercial background agents, I would call them, that you can buy off the shelf work very well for web, like standard web apps. I'm very happy with them. If you are not using one of them as an engineering organization, maybe it's like doing classic SaaS. Highly, highly, highly recommend. But I am hearing more and more from teams to things that you called out. I am hearing more and more people are just building their own background agents. So it's happening, it's happening in teams. Very, very big and very, very small. I think the primitives are there to start experimenting with it. And so I don't think it's going to be as surprising to us to hear about people building their own internal coding background agents. Even if like core infrastructure is something from the big, the big models, model providers. I think the second thing that I'm hearing a lot, and we heard this from the Stripe team is investment in cloud development environments and remote computing. Again, because if you were to run some of the stuff, especially the data heavy stuff on your computer, it starts to sound like an airplane taking off. It's no good. And then the last thing I heard you say, which is like, ports, I joke with everybody. I say work trees everywhere. Ports 3000 through 3009 accounted for like, I am just like every everything. And I have to call out. Chris Tate at Vercel released a thing called Portless, which just makes managing multiple ports local host ports on your local machine a little nicer. So for simple things, I would go look that up. We'll link it in the, in the GitHub show notes. But you know, common problems that I think people have running concurrent engineering processes on their own machine. And then the like, meta thing, which is just like, make time to code.

[B]

You need it, everyone.

[A]

I also don't take meetings after one. Sometimes I'll do podcasts in the early afternoon for folks, but all afternoon I'm just like in my real state, which is hoodie on bad products.

[B]

Yeah. And I think that I'm sure you feel this too. But like, when I was handwriting most of my code, I would enter this sort of like euphoric flow state where I, you know, just completely focused on a problem. And then when I started doing a lot of agent coding, I lost that for a little bit. But now when I'm writing code, you know, Lane 8 just released a new album yesterday. You should listen to it. Put on your, your hood and your, your headphones. I'm like, way. I'm like totally back in that state now. Just doing a different workflow.

[A]

Yeah. And I'll give folks this sort of, you know, AI mom of the Internet that I try to be, which is. I do feel like a lot of people are, they, they kind of go into two camps. They are having more fun than they've ever had before, and they're back in the flow state of like, what got them into software engineering or building or technology or whatever. Or they're approaching like clot anxiety, burnout breakdown because they feel this like, productivity anxiety. And they're not. I, I think, I think what I see is that people feel like if they're in a meeting and they're not kicking off agents, they're doing something wrong or if they're talking to somebody and they're not kicking off agents or doing. And I just say, like, I like the idea of chunking your, chunking your time with AI a little bit more Yeah, I think it just narrows you on the more. More productive pieces of it and is also just a more enjoyable way to get stuff done.

[B]

Yeah, I had a phase, which I think I'm over, you know, my wife Elena, where we would have. We all. We have dinner together usually like pretty much every night. And so I had a phase where my laptop was not at the table, but open and on the couch. And I think I've progressed beyond that phase now. So now the laptop is closed and it. I think it's. It's an important. It's an important thing.

[A]

I. I agree. I. When I was first using Openclaw, I installed it on an old MacBook and it would, like, stay open on our kitchen island, which is where all our plugs are, and it would, like, hover over us at dinner and hover over us at. At breakfast. And if it got moved, I was like, where is Polly? Is she alive? Is she open? Is she closed? So yes. Close your laptop, people. Close your laptop. This episode is brought to you by Persona. You're learning to build with AI, but there's an important question you need to ask. Who is actually using your product? Is it a legitimate user, a bot, or a fraudster? Brex, Figma, Etsy, and Twilio trust Persona to answer that question. With Persona's identity verification platform, you can create branded experience experiences, automate fraud prevention, and know who is human online. That makes it easy to give good users an experience that makes them feel welcome and to stop bad actors from causing damage. And for those of you building in the AI agent space, Persona helps you verify the identities of people, businesses and developers behind agents. It's how companies like Lithic and Skyfire are pushing the frontier of agentic commerce. Learn more at with Persona. Com. All right, so, you know, we covered the first half of this episode, which I think is very interesting for technical folks. How to have kind of like long running or just really diligent agents run against technical problems to give you real benchmarks about performance on changing things. I love that. Second thing is just your core workflow on how you do coding, both how you dedicate time, and then technically just what your workflow looks like. Let's talk about evals, because I feel like this is something that's very intimidating to a lot of people and obviously you build a product that supports this. But taking a step back, why do you think this concept is so important to understand? How can you just demystify it for folks who are a little intimidated by it?

[B]

Machine learning specifically shifts the Task of programming from being about the how to being about the what. And this is true, like, forget about LLMs. Like, you know, it's true with, let's say, like, you're back in, like, middle school. You're doing, like, remember, statistical regression. You're not defining the. You're. You're computing what the slope and the Y intercept should be. You're not defining it, but you give it all the points, which are the, you know, the what, not the how, which is the slope and the Y intercept. And I think that, you know, the cool innovation around, like, transformers and the next token prediction task, which lets you, you know, ablate tokens and do all this cool stuff, it's all about saying, like, okay, here's like, the compute substrate and here's the what, which is the outcome. It's predicting the next token. Can you go and use a lot of GPUs and figure out how to achieve that? And I think that if you take that as inspiration for anything you do with AI, then you're able to be more productive. And I think that applies to traditional programming, like what we just talked about. I'm not dictating exactly the implementation or even the set of algorithms that we're using to solve problems. I'm just trying to define very succinctly what the problem is and why it is a problem and how to assess the solutions to the problem. It also applies to building AI software. And that's what evals are all about. Evals are a methodology for you to say, this is what success looks like. In my opinion, evals are actually the modern version of a prd. So a prd, you would say, hey, in pros, this is what success looks like. Evals are also often written in prose, but you supplement that with examples. So, you know, the best PRDs, they have good examples. Like, maybe someone's made a demo or written out, like a user story or something. It's the same thing. The difference with evals is you encode those user stories in a way that can be quantified to some extent, and then you let a model or whatever figure out the how, and you are really focused on the what.

[A]

Give me an example of how you use this in product development just to make it a little bit more tangible for folks.

[B]

Yeah, let's start with something that I think is quite straightforward, and then we can venture into the less straightforward stuff as we go. So this is our ui. And like, I'm working on a very simple task here, which is I'm trying to Create a prompt that will be part of an agent that is good at answering questions about Brain Trust documentation. So we looked at a few questions that people are asking in our docs and we just put them into a data set. You can like upload a CSV file like it doesn't matter, it's just come up with a list of some questions or you can auto generate them, whatever, just start somewhere and wrote like a very basic prompt. We're going to use GPT5.4 mini. And I attached an MCP server. So I attached the BrainTrust MCP server. We were also playing around with Context 7, which indexes docs for you. You could also turn off the MCP and just see what the model already knows about your product. They're getting pretty good at knowing about every product now as well. And here I just, I just ran it and so you can see some of the answers. I'm going to be honest though, I don't really want to read all of these manually. And so what I would usually do is I just start by saying like, hey, can you come up with a good scoring function for these outputs? I care about having concise code snippets only using one language and let's say avoiding EM dashes

[A]

always.

[B]

Yeah, of course. And so now in this case, GPT 5.4 is going to go and actually like look at all this stuff for me and it's going to look at some of the outputs and it's going to rerun stuff and it'll kind of do its thing and it's going to come up with a new scoring function. One of the things I think, by the way, that's kind of cool about this workflow in general, and I expect to see this in more products over time, is that you'll notice like I have this in the equivalent of like unhinged mode of a coding agent, which is sometimes dangerous to run on your machine. But this agent is running inside of this playground and it's using like data and some prompts and stuff. So the risk of letting it just go and try stuff out is actually very low. And so I think I'm excited just generally about seeing agents in more environments outside of my local computer with BASH and something that's like very dangerous and could screw up my life if it goes wrong. I'm excited about just having more agents that sort of run in these types of environments and do whatever they want. Like, I don't even know what this is doing right now, but we'll find out in a Few minutes.

[A]

I'm really excited about this. And just for people that are not watching or need just another set of context, basically what you did is you took these questions, questions that people are asking in your doc site or search or whatever chat bot about how the product worked. You built a little prompt to answer those questions. And then right now you're building. You're having AI build a score that tells you how well these questions are getting answered based on like a very loose definition of what you want it to do. And then is that score apply that scoring mechanism applied across all of these so you can actually rank it?

[B]

Yes. Yeah. Yeah. I think it's going a little bit awry, actually. So I'm going to switch to this one, which is a little bit better.

[A]

Do it live. We love a live demo.

[B]

I know. And let's use. Let's use Claude and give it a shot. So this one is a little bit cleaner. And it actually wrote a prompt. Well, let's use a smarter model. It didn't pick the smartest model. It wrote a prompt which takes the input and the output and then it evaluates it on these criteria. It is a pain in the ass to write these criteria out by hand. So it's really nice to just let a model do it for you. And what we can do is run it and it will quantify how well the model does on these criteria and then we can look at it one by one. Or what I actually tend to do nowadays is look at it in aggregate. And so the scores will start coming in here.

[A]

What's the alternative that people are. What do you see people doing as an alternative? This that you think is less effective one is just not doing it. I know a lot of people.

[B]

Yeah, I mean, I think that a lot of people. And I fall into this trap myself despite working on this product. So it's not. There's no judgment for. For doing this. But I think what a lot of people do is they just try stuff out on one or two examples and they try to generalize from that. And frankly, I don't think that's a bad idea. Like, I think that vibe checks are extremely important. But what's. What happens is that if you do this, you end up playing kind of like a whack a mole game. So you might make it really good at one or two things, then you ship it, and then it's not good at something else. And what we do, we have this designer named David. And David is really cool. Like, he dresses well, he has like. Like he's into the latest music. He. He, like, he likes music before other people do. He told me that when he was a kid, he. He played soccer and everyone had black shoes and he wanted the orange ones, and then the next year, everyone wanted the orange shoes. So he's like that kind of person, right?

[A]

Yeah.

[B]

And we have a lot of AI stuff going on, so it's not practical for David, who has, like, the ultimate. Who's the ultimate brain trust tastemaker, to look at everything manually. And what I actually do is I run a shit ton of evals to try to quantitatively improve things. And then when I feel like the evals are good and my own less sophisticated palette thinks that the results are good, I will go to David and ask him for a vibe check. And I probably do that, like, once every few days. And then David gives me the vibe check. And like, half the time he just completely destroys everything that I've said. Like, hey, you know, you think it's good, but it's actually not very good. And then what happens is I will go back and try to capture what David said, and I'll say, like, you know, hey, David actually thinks it's okay to show both languages as long as, you know, blah, blah, blah, blah, blah. And then I will. So I'll try to sort of capture David and then improve the scores and then attempt to quantify David. And then the next time I go to him, I don't, like, repeat the same mistake, but I still get his vibe check.

[A]

Well, and I just have to call out the. The meta thing here in this David story, which I love, which is I have a lot of people saying, wow, if I go as so far as to turn my own taste or my own skills or own expertise into a system, whether that system is like the David eval, the David David in a loop judge, or something else, I. I'm. I'm functionally just building my own replacement. And I am presuming, because I do, and it sounds like you do too, you value David more in this system.

[B]

Oh, yeah, yeah, yeah. We're able to have David's pallet applied to more things like the. The. I think the quality bar is that we're able to hit is higher because we are able to get more things to that. To that bar.

[A]

I love it. Okay, so this has been a powerhouse episode, one of my favorites. We've talked about a lot about, you know, solving really technical problems with AI. We've demystified evals a little bit for folks and shown how in. In a Safe space, you can actually let AI. I think that's one of the meta themes of this is in a safe space you can let AI run with a lot of autonomy and you'll throw a lot of data at it. And you can get higher quality outcomes, much more so than if you were to manually fix things or even manually evaluate things. I'm going to do a quick lightning round and then we'll get you back to. I mean, it's almost noon, so back.

[B]

It's time to code.

[A]

Time to code. One, I have a question. When you say there is no excuse, there's no excuse for bugs, there's no excuse for little design knits, there's no excuse for that. How do you feel like you practically. I maybe have two questions that you can answer. They'll be our two laying around. How do you practically manage the velocity to customers, which is, do you ever get customers being like, wait, what's this? Wait, what's that like too much features just consumed as a customer. And then two, how do you technically manage the throughput into the system?

[B]

Product building and code writing is now looks like carving rather than constructing. So it's very fast to create something that has too many features and too many buttons and too much code. And you need to spend a lot of time removing stuff. And so we actually, I would say 90% of the time someone complains about something, we remove the thing that was causing confusion and just make the system work better because we understand now that the person complained, their point of view. And we're able to build a product that doesn't even need the complexity that led them to the confusion in the first place. I'll give you an example. If you load a trace and you imagine hitting command F, you might in your brain think that that's just searching what's on the page, but what's on the page might be hundreds of megabytes of text and it's virtualized and it's across spans and there's also a table. So we had a very powerful search implementation that would search across the spans and rank everything and you know, blah, blah, all this cool stuff. And then a lot of people complained and they were just like, why is this? You know, I just hit command F. I just want it to show the thing. And we've just, we've, we've really simplified it over time. So I think, I think we try to carve. And then in terms of technically managing it, we spend a lot more time working on CI than we used to. And so I Think that a lot of platforms effort has shifted so that if we are really good at CI then we are able to move faster. And if we feel like we're constrained, then instead of shipping a bunch of crappy stuff, we're like, okay, let's pause and improve CI so that we earn the ability to move faster.

[A]

Okay again for the VP of engineering in the back, invest in C. I've, I've told everybody, they're like, how do I accelerate my engineering velocity with AI? I was like, fix your CI.

[B]

Yeah, yeah. I mean I think every engineer is now building a platform and upon the platform agents are doing the work that the engineers were doing manually. Right. And I think that applies to evals. Like if you're an engineering team and you're building an AI product, the number one job for you is to build a feedback loop, meaning you have a pipeline that allows you to summon from the ether of real world data and turn that into evals. And as an engineering team that is your number one job. It is not prompt engineering, it's not picking an agent framework, it's not rewriting your database, whatever. It's creating that pipeline. And the same is true. CI is that same idea, but applied to software engineering.

[A]

Well, and I'll give one other tip which is you think that those evals people are always like, oh yeah, for my AI product, I need that. I have seen again, I think the intercom team has run a bunch of evals on their internal use of quad code to figure out where engineers are hitting pain points, where people are giving up, where the agents are asking for permissions that have to be escalated. And I think that sort of analysis on your team is very, very important and ultimately gets you to these, these better outcomes. Okay, last question. You seem like a very reasoned person, so I'm presuming I'm going to get a very reasonable answer. But I ask everybody when a one of one of your four tabs is not doing what you want, when the evals are failing the David test, what is in your back pocket prompting strategy that you, that, that you rely on? Do you yell, do you bribe?

[B]

Close the session. And then I improve the evals and then I try from scratch again. Yeah, yeah.

[A]

This is a man who is on message.

[B]

Yeah, I mean I'll give you like an example. We have this open source use case, sorry, a use case where we run open source models and we're running like millions of tokens per second. It's very, very high scale. So every cent matters and every bit of Optimization matters. We are trying to change right now from model A to model B. And again, I'm someone who builds software to write evals. I vibe coded an eval script and it went. It just was getting stuck. And then I read the code and it's like 3,000 lines of complete trash. And it had like all these scoring functions and all this crap and it was getting confused. And so I. On Saturday I hand wrote like, no, no copilot, no autocomplete. I just, I. Partly to improve my own understanding of the problem, I hand wrote the eval and then by the end of Sunday, the problem was solved. Yeah.

[A]

So you shut the session. You do it yourself.

[B]

Yeah. Just for the evals. Just for the eval.

[A]

Great. This has been so great. Where can we. Where can we find you and how can we be helpful?

[B]

If you are interested in evals or you're trying to solve AI observability problems inside of your company, please check out BrainTrust. We're at braintrust.dev@braintrust on X. Or I'm at A N, K, R, G, Y L. Very happy to chat. We're also hiring. If you like working on these problems and you like maybe pushing the boundaries of rigor and stuff and found this kind of stuff interesting, we'd love to work with you.

[A]

Well, thank you so much for joining. This was great.

[B]

It was a lot of fun.

[A]

Thanks so much for watching. If you enjoyed this show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify or your favorite podcast app. Please consider leaving us a rating and review which will help others find the show. You can see all our episodes and learn more about the [e-mail] See you next time.