The maturity phases of running evals — Phil Hetzel, Braintrust

Hinzugefügt: 2026-05-28

[00:00:14]Welcome everyone. Um, it's always a challenge to be a presenter directly after lunch because that's typically when the energy level goes from right around here to around here, but I'm going to try to make this session worth your while uh today. We've got 18 very quick minutes uh together and uh during that time I'm going to be talking about the uh different maturity levels that I see people go through as they perform evals for their agents.

[00:00:41]Uh before we get into that, um just roughly quick agenda today, um I'll explain a little bit about myself, the company that I work for. We'll spend most of the time today on on more theoretical concepts, not product concepts. And then um we'll we'll talk about um where I I think this field is going in the future. Uh I'll also make sure to leave enough time, hopefully a couple minutes, for questions as well. I didn't over prepare the content and in hopes that we could have a little bit more of a discussion uh at the end of this.

[00:01:11]Uh first of all, this is me. My name is Phil Hetzel. I lead solutions engineering for a company called BrainTrust. Uh effectively what that means is that it is me and my team's job to make sure that people are getting the most value out of the platform as quickly as possible. Prior to BrainTrust, I spent 12 years in consulting and systems implementation.

[00:01:33]Uh first 4 years with KPMG, last uh 8 years in consulting with a company called Slalom Consulting. And with Slalom, I led their global Databricks business unit. And I noticed that a lot of my customers were prolific at creating generative AI proofs of concepts. They were not as prolific at bringing those proofs of concepts to production. So, I started using BrainTrust first as a user because I wanted to help bridge that gap uh for my customers. And I like the product so much that I ended up joining the company and I've I've been here for about a year.

[00:02:07]Um outside of of work, I like to play chess, but I'm not very good at it. And I like to spend time with my wife and my dachshund. His name's Pistol Pete. Um he's the one in brown, not the one in in black.

[00:02:20]Uh what is Braintrust? Uh the company that I work for. Braintrust is an agent quality company.

[00:02:26]Um one of the I guess two of the main ways that we contribute to agent quality are evals and observability, which we consider to be very much the same problem uh from a systems perspective.

[00:02:39]Evals, of course, being the thing that you're doing in order to gain confidence in your agent as you want to bring it to production. And then observability being the practice of once that agent is in production, remain in confident in it.

[00:02:55]Um it's a growing space, it's a very fast-moving space. And uh uh in the uh when you build an evals platform, you really have to grow with the technology, the underlying technology as it changes.

[00:03:06]So, it's a very fun uh fun place to be in.

[00:03:10]Uh let me give like a a a quick overview of the problem. Uh we talked a little bit about why we do evals in the first place. How many of you of you all are doing evals today? Hopefully, as as you build, um every single hand should be up. And and certainly, when I give this talk next year at this conference, all you're going to come back, of course, to this session and every hand is going to be up. Eval is very important. The reason why we do evals is wholly in service to agent quality. That's the most important thing. We want to make sure that our agents are doing what we expect when confronted with real usage and and and real users.

[00:03:47]Um this is really important from a risk perspective and and a brand perspective.

[00:03:52]We don't want um the reputational risk of an agent being unkind or unhelpful to a customer. We don't want the systems risk of an agent costing us too much money as it as it operates.

[00:04:05]And there could even be compliance and legal risks if your agent goes too far off the rails. So evals are a both a defense against those types of risks.

[00:04:14]But they're also they can play offense with evals in knowing with each tweak that you make to your agent how it's improving and how much it's improving your application.

[00:04:26]A couple of primitives here. Evals are not unit tests where whereas unit tests are very exhaustive in in how you perform them. With evals you want to make sure that you start very high-level with the failure modes of your agent.

[00:04:41]Either either you or a subject matter expert can educate about the specific failure modes of an agent and you build evals around those very specifically.

[00:04:51]What you don't do like you would with unit test is think about exhaustively every single thing that could potentially go wrong with your agent and try to make an eval for it. Why can't we do that? Because it's it's infinite. You would spend all of your time writing tests and none of your time shipping, which is which is not productive.

[00:05:07]Eval results don't need to be perfect.

[00:05:10]Sometimes they can be Sometimes they can be directional.

[00:05:13]Using LLMs to judge other LLMs LLM as judge techniques, you're probably not going to get 100% every time. That's okay. As long as you're trending in the right directions with those more non-deterministic techniques, that that is completely fine.

[00:05:28]Um Different primitives with the eval itself how it's constructed. You have three things. You have a task. That's the agent under test or the prompt under test.

[00:05:39]You have some data set of examples that initiate that task. How do you invoke that task? You use some example that you give to an LLM or give to an agent to to start that workflow and then you have certain scoring functions which you're using to judge the utility or the quality of that task.

[00:05:59]Um there are a couple of different maturity areas that I've noticed some of our customers go through. I've listed four here. This is probably more of a continuum than than being very discreet.

[00:06:10]But suffice to say that these stages you know you will you will traverse these stages as and when you create more complexity within your agent just by just by necessity.

[00:06:23]The more complex agent you're that you're building the more vectors there are for failure the more failure modes you that you may need to account for.

[00:06:33]Um we're only going to be focusing on the like eval theory itself today. We're not going going to really talk about the platform surrounding evals. We've got a booth for that downstairs. If if you're interested you can come find me.

[00:06:47]So go through these four just getting started measuring to manage accounting for complexity and then some advanced eval techniques.

[00:06:57]Okay, just getting started. It's not wrong to just get started with with vibes. I know like like vibe checking is a very nasty phrase here at this conference. I actually think it's okay. It's it's certainly better than nothing.

[00:07:11]When you're first starting out you can't help but start with vibes. I think the only thing that I would really recommend is that as you are vibe checking you're also documenting.

[00:07:21]So when you have an agent under test you give that agent maybe 10 different example inputs and and and loop through those inputs to see what the output is.

[00:07:32]You should probably have some human whether it's the person who built the agent or even better a subject matter expert that really knows what a quality response would look like. You should really have them analyze these outputs and and give two pieces of information.

[00:07:49]You should give a thumbs up or thumbs down. Is it Was this response good? Was it bad?

[00:07:55]But more importantly, you should make that human annotator um perform a justification for why they chose that thumbs up up or thumbs down.

[00:08:04]Reason being is that you you're you need to extract a lot of this domain-specific knowledge out of that human annotator's head so that eventually you can scale that type of knowledge through a through a technique like like LMS judge. But it's a great first step performing human human annotation. Who who in this room is like at at at this step?

[00:08:26]Well, okay, this is way more advanced group. That's okay. That's okay. That's a good place to start. Are you using like human expert and annotators? I I just got like a I mean, I I have plenty of structure set up. Yeah.

[00:08:39]Uh yeah, it's pure running agent on my data.

[00:08:44]You guys see how it looks. Yeah.

[00:08:47]Yeah, totally. That's it's it's it's You have to start somewhere. It's a great place to start. Um Uh this is this is how like that that workflow is going to is going to look.

[00:08:55]You have a trace come in, thumbs up thumbs up or thumbs down, and then you add some justification um to that so that eventually you can use it as as an LMS judge score down the line.

[00:09:09]Um this is like what this might look like in a in a platform like Braintrust.

[00:09:14]Um Uh we have like a like a human annotator view uh built in the platform. We actually let you live code your own annotation views.

[00:09:22]Um important point, don't give a generic uh annotation platform to users. Really make it very specific to them. They're going to have an idea of how these agent traces should look. So, you should deliver that to them and it'll encourage them to um evaluate these these appropriately.

[00:09:43]Um okay, the next the next part is expanding upon that a bit where now I just don't I don't have only some human greater giving thumbs up and and thumbs down and justification. Now I'm starting to use those justifications and I'm I'm probably running those justifications through cursor cloud code or or codex to try to derive the actual failure modes of why when they gave a thumbs down why they delivered a thumbs down. The these you're you now know and understand the failure modes of your agent.

[00:10:17]Now that you understand the failure modes of your agent, you want to be able to scale that human knowledge and be able to to automate it so that you're not dependent on a few people with expertise to judge a agent agent outputs.

[00:10:31]A couple ways to a couple ways to do this, one of which using LLMs to judge other LLMs. LLMs judge we we've that concept's been around for for quite some time. Very effective. Um important here is that whenever you use an LLM judge, just because you put a robe and a cloak on an LLM, that doesn't make it inherently more trustworthy. You should be evaluating LLMs as judge outputs as well.

[00:10:57]Um that's not really covered in this presentation, but um you should not just judge LLM judges blindly in that regard.

[00:11:04]Uh there also might be some objective failure modes where you can deterministically um encounter them just through code. That's okay, too. You don't have to use LLMs to judge other LLMs. You can use code to understand um if you're using too many tool calls, you might want to fail that eval as an example. If we're using too many tokens, you might want to fail that eval.

[00:11:27]Um I think the most important point here is that this data set uh, that's that's on the right-hand side of this slide. At this point, you should probably be gathering production traces or at least, uh, UAT level traces into that evaluation data set. We want it to be very like don't think about evals as running tests. Think about evals like rerunning production because ultimately we want to be confident as we run, uh, run these workloads in in production.

[00:11:58]Great way to do that is just to capture production data.

[00:12:02]Um, most important point is is this. Uh, we we call it like the the flywheel internally. We want to be able to capture these traces these agent traces in production, understand what's going wrong with them either through a human or or through automated tooling, um, and then bring those examples back to some offline experimentation environment, rerun production through an eval, and then use that to guide us to which direction we should be improving our agent. So, evals that that's like more playing offense with with your evals.

[00:12:39]Um, this is just like a an example of, uh, setting up an setting up an LLM as a as judge scoring function to expand, um, your your ability to evaluate at scale rather than using just a human.

[00:12:54]Um, okay, uh, level two. Now, we're starting to not just do simple model calls. We might be performing work with external systems. I think of of tool calls in two different ways. There is context gathering tools that are just, uh, gathering data and injecting that into the LLM. And then there is CRUD-based tools where you're creating, reading, updating, or deleting information from a database or an external system. Um, both of these are, uh, can have a lot of lift in terms of whether your agent is quality or not. It also means that there's a lot of other things that can go wrong with your agent when you're starting to interact with external external systems.

[00:13:37]Often now instead of just having one um of evaluating one specific part, i.e. the output of an agent, now you might be having to evaluate the entire trace of an agent. So in that sense like this is where tooling starts to come into play.

[00:13:56]You'll you'll need some way to capture these large traces, understand each and every step that an agent took to be able to introspect and eventually target evals towards maybe even individual tool or MCP calls that you're that your agent is creating.

[00:14:14]The other the other problem here that that we might have is when you're performing CRUD on on a system, you you don't really want to do that when you're offline. Of course, there might not be a way to do that when you're offline. So when you run an eval um there's there's two things that are problem areas. One, really challenging to represent the state that the other external systems were were in at the time that eval input was created.

[00:14:44]And then two, it makes it really challenging to interact with those systems that the agent could be interacting with because you don't want to overwrite any production data. These are real challenges that we have to solve for. I would say it's not completely solved right now.

[00:14:59]Um however, um there there does need there there are some ways where you can represent external system state and interact with like mock level APIs so that you can approximate real a real production environment in as you're running evals.

[00:15:18]Um the idea for this is that um these these agent traces can be arbitrarily large. Um in that sense it's it's a lot different than application tracing.

[00:15:29]So, if a trace can be arbitrarily large, you can actually cram in a ton of context, i.e. system state, the state that the external systems were in at the time, into these traces and inject that into a into the task that you're running the eval upon.

[00:15:46]In that way, uh instead of having to uh um uh create an entire test structures and and and infrastructure, you can represent a lot of that stuff within the trace itself and then encapsulate it there.

[00:16:01]The other thing that you can do is you can use like really really specific querying techniques to um perform timestamp queries uh to systems that support them. So, if an input came um and and you added it to your data set at a certain point in time, perhaps uh the way that you've set up your vector database, you can run a version query to query the vector database at a certain point in time. So, that way you're adequately representing the state of uh of of when that task ran originally.

[00:16:35]These are more complex techniques, um but ones that ones that are that are a little bit more emerging.

[00:16:42]Um I only have about 2 minutes left to go.

[00:16:47]Um what's next? Um performing topic modeling at scale to make sure that you're uncovering those failure modes automatically in production. Um that's something that like more than happy to talk about at uh at the booth downstairs. And then of course, performing evals in a way where you're um using cloud code and the eval provider CLI to be able to do this in an automated automated way.

[00:17:11]These are two other patterns that uh that I see emerging in the space.

[00:17:15]I want to be conscious of time. I probably have a for like one question uh before I have to jump here. Is anyone curious about anything specifically?

[00:17:22]Otherwise, you can find me at the booth.

[00:17:24]Yes, sir. In our sphere, like it's kind of normal to put a bit more respect on deterministic evaluation and deterministic graders. Do you agree with it? Do you think that we should push for more deterministic graders in your evaluation platforms or do we embrace LLMs as judge as a result?

[00:17:45]I would Some things are subjective.

[00:17:47]That's why we love agents so much. I would embrace LLM as judge, but also perform a lot of evals on the LLM as judge so that like it's it's very aligned with what a human would decide in the same circumstance. You would eval the eval as an eval. Yeah, it's easier to do that because LLM judge outputs are going to are going to be discrete. So, you can create a ground truth data set for that. Yeah.

[00:18:09]All right, everyone. I have to jump. I'm at my time. It was a pleasure to be with you all. And yeah, feel free to find me in the booth downstairs.