AI Engineer World's Fair 2025
The Future of Evals - Ankur Goyal, Braintrust
Read the talk
From Eval Dashboards to Agents That Improve Evals
Braintrust’s Loop moves evaluation beyond inspecting scores, using frontier models to propose changes to prompts, datasets, and scorers while keeping those edits visible.
From a talk by Ankur Goyal
The dashboard still leaves the hard decision to you
After an eval finishes, what should you change in your code or prompt? Running evaluations can become routine long before answering that question does. After almost two years working with AI product teams at Braintrust, Ankur Goyal describes substantial evaluation activity: Goyal reports that the average organization signing up for Braintrust runs almost 13 evals a day, while some customers run more than 3,000 a day. Some advanced customers also spend more than two hours a day in the product working through evaluations. These are reported product-usage figures; the observation window and customer cohorts are unspecified.
Even teams building highly automated agents still do much of this improvement work manually. An eval produces a dashboard; a person inspects it, leaves the dashboard, and decides which code or prompt changes might improve the next result. The missing step is turning evaluation results into useful changes. A better dashboard helps people inspect the evidence, but it still leaves that decision with them.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Evaluating models that improve evaluations
Loop is an agent built into Braintrust to help perform that improvement work. Its development depended on evaluations of the models doing the improving: every quarter for two years, Braintrust tested frontier models on their ability to improve prompts, datasets, and scorers. Goyal says those models performed poorly until very recently.
Goyal reports that Claude 4 performed almost six times better than the previous leading model in Braintrust’s internal tests of evaluation-development tasks. The talk does not specify the scoring metric, test-set composition, inference settings, previous leader, or which Claude 4 variant was used. The comparison explains Braintrust’s decision to pursue this workflow, but does not provide enough detail to reproduce the result.
Loop’s scope extends from optimizing prompts to improving complex agents. It also helps build datasets and scorers, because evaluation quality depends on all three working together.
| Component | What Loop helps improve |
|---|---|
| Prompts and agents | The behavior being evaluated |
| Datasets | The cases used to evaluate that behavior |
| Scorers | The criteria used to judge the results |
This makes the evaluation setup itself part of the work: an agent can help improve how a system is tested as well as the system under test.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Enabling Loop and choosing a model
The launch-era setup presented in the recording is short:
- Sign in to Braintrust or create an account.
- Enable the feature flag named
Loop. - Use the default Claude 4 model or select another model you have access to.
Goyal names OpenAI models, Gemini, and users’ own LLMs as alternatives. These are the instructions and model choices announced at the talk; today’s Loop documentation describes access within the product and model selection through providers configured for the organization.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Keep proposed changes visible
Loop runs directly inside Braintrust, preserving the ability to inspect data and prompts while working on them. That matters because an optimization suggestion may change different parts of the evaluation. The interface shows proposed changes side by side, including:
- Data edits: changes to the examples being evaluated.
- Scoring ideas: suggestions for how to assess outputs.
- Prompt edits: changes to the instructions driving the system.
The proposal remains connected to the material it would change, so users can examine the edit directly in the working interface.
For users who want more autonomy, Goyal describes an optional toggle as “Just go for it”: Loop proceeds with optimization rather than stopping at suggestions for inspection. He says that mode works well, without providing a separate measurement of its effectiveness. The product therefore offers both visible proposed edits and an option to let the agent optimize autonomously.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Evaluation becomes work for the agent, too
Evals have become a critical part of building AI products, while the work of interpreting and improving them remains labor intensive. Speaking at the June 2025 AI Engineer World’s Fair, Goyal forecasts that frontier models will transform evaluation itself over the following year. Braintrust’s intended direction is to incorporate those advances into the tools used to develop and evaluate AI systems.
That direction is still being shaped. Goyal invites users to try Loop and provide feedback, acknowledging that substantial work remains. He also invites people working on UI, AI, and infrastructure to join the team, closing with an on-screen QR code as a route to get in touch.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Anthropic's original announcement of Claude Opus 4 and Sonnet 4, including capabilities and benchmark methodology.
Further reading
A recap of Goyal's technical Q&A covering small initial datasets, scoring functions, and iterative evaluation.
Updates since the talk
Instructions for using Loop to analyze traces, optimize prompts, generate datasets and scorers, select models, and configure edit approval.
A November 2025 account of Loop's production-data analysis workflows, team use cases, and product access.
Read the complete timestamped transcript
- 0:00
[upbeat music] [audience applauding] Awesome.
- 0:22
Uh, so today we're gonna talk a little bit about evals to date and where we think evals are gonna be going in the future.
- 0:32
Also, for those of you who saw my brother earlier, um, I'm gonna do my best to live up to his energy and, uh, and charisma. [laughs]
- 0:42
But, um, yeah, you know, it's been an amazing almost two-year journey for us at Braintrust. We have had the opportunity to work with some of the most amazing companies building, um, I think, the best AI products in the world.
- 0:55
Uh, I'm blown away by how many evals people actually run on the product. The average org that signs up for Braintrust runs almost thirteen evals a day. Some of our customers run more than three thousand evals a day.
- 1:11
Uh, and some of the most advanced companies that are running evals are spending more than two hours in the product every day working through their evals. And I think one of the things that stands out to me is while we have customers building some of the coolest, most automated, um, AI-based products and agents in the world, evals
- 1:33
are such a manual process. Um, you know, to date, every time you run an eval, the best thing you can do is look at a dashboard, and I think we have a pretty cool dashboard in Braintrust.
- 1:44
But still, it's just a dashboard that you look at and you walk away and think, "Okay, what changes can I make to my code or to my prompts so that this eval does better?"
- 1:55
Um, and I actually think that is all going to change.
- 2:00
Uh, so today I'm excited to talk about something called Loop. Loop is an agent that we've been working on for some time now that's built into Braintrust, um, and it's actually only possible because of evals.
- 2:12
Every quarter for the last two years, we've run evals on the frontier models to see how good they are at actually improving prompts, improving datasets, and improving scorers. And until very, very recently, they actually weren't very good.
- 2:27
In fact, we think that Claude 4, in particular, was a real breakthrough moment, um, and it performs almost six times better than the l- the previous leading model before it.
- 2:39
So Loop runs inside of Braintrust, and it can automatically optimize, uh, your prompts all the way to very complex, uh, agents. Um, but just as importantly, it also helps you build better datasets and better scorers because it's really the combination of these three things that make for really great evals.
- 3:00
This is a little preview of, of the UI. Um, you can actually start using it today if you are an existing Braintrust user or you sign up for the product.
- 3:08
There's a feature flag that you can just flip on called Loop and start using it right away. Um, by default, it uses Claude 4, but you can actually pick any model that you have access to and start using it, whether it's an OpenAI model, a Gemini model, or maybe some of you are building your own LLMs.
- 3:24
You can use those as well. Um, and as you can see, it runs directly inside of Braintrust. One of the things that we, uh, learned from working with a lot of users is how important it is to actually look at data and look at prompts while you're working with them, and we didn't want that to go away,
- 3:43
uh, when we introduced Loop. So every time it suggests an edit to your data, or it suggests a new idea for scoring, or it suggests an edit to one of your prompts, you can actually see that side by side directly in the UI.
- 3:56
Um, of course, for the more adventurous among you, there's also a toggle that you can turn on that says, like, "Just go for it," and it will go and optimize away, um, which actually works really well.
- 4:10
So just to recap, uh, to date, evals have been a critical part of building some of the best AI products in the world, but the task of actually doing evaluation has been incredibly manual.
- 4:23
And I'm excited about how over the next year, uh, evals themselves are going to be completely revolutionized by the latest and greatest that's coming out, um, from, you know, the frontier models themselves.
- 4:34
And we're very excited to incorporate that into Braintrust. Please, if you're not already using the product, try it out. Uh, try out Loop. Give us your feedback. Uh, we have a lot of work to do, um, and we'd love to talk to you.
- 4:46
We're also hiring, uh, so if you're interested in working on this kind of problem, whether it's the UI part of it, the AI part of it, or the infrastructure, uh, side of it, we'd love to talk to you.
- 4:56
Um, you can scan this QR code. Uh, it should be over there. Yeah. You can scan the QR code and, and get in touch with us. Uh, we'd love to chat.
- 5:04
Thank you. [audience applauding] [upbeat music]