We built Brainstore for the amount of data agents generate. Now agents are helping teams search through traces and run far more queries than a person would. Nitro is Brainstore's new asynchronous query execution engine. It separates waiting for storage from processing the data it returns, bypassing the limits of traditional object storage. We ran Nitro on more than 1 million real queries and found that full-text searches were over 2x faster. It's already on for Braintrust SaaS and BYOC customers. Read more →

We tested where Jev holds up as a judge. Jev was fast, cheap, and highly competitive for judging groundedness. But it lagged behind models with reasoning for math and code domains. Use it for certain eval tasks, but don't throw out your LLM-as-a-judge just yet. Read more →

Jev from @typesafeai can replace your LLM-as-a-judge for scoring agent responses. It returns a choice or numeric result with information about uncertainty, so you don't need to spend time and resources prompting a general-purpose model into an LLM judge. Use Jev as a judge scorer in Braintrust and review its selected answer, confidence, and probabilities alongside the score. Trace Jev calls from your own application with the JavaScript or Python SDK. Read more →

What's new: - Surface hidden insights in your traces with Patterns - Schedule custom recurring analysis of your traces with Loop automations - Keep multiple human reviews independent with blind reviews - Manage access to your org-level AI providers with fine-grained permissions - Score your multi-turn traces as one unit with group-scoped scores - Search, star, and section your dashboards Read more →

When you eval open-weight models, you also need to eval the serving stack. The inference engine, model revision, precision, caching, and account limits can all impact the system you're testing. We ran an eval on Kimi K3 served on @FireworksAI_HQ vs @Kimi_Moonshot. They performed equally well on quality, but median time to first token was 55% faster on Fireworks. Read more →

Engineers at @evelegalai build legal agents that summarize depositions, conduct research across case law, and draft work product. Patterns proactively finds the issues affecting only 1% of agent outputs, and helps resolve these silent regressions. Try Patterns →

GPT-6 Astra is now available in Loop. Use it for pattern detection, eval creation, trace debugging, and other production investigations. Loop is built on Codex, so you get the same model capabilities on top of all your Braintrust context.

The Braintrust MCP server now exposes write tools. Your coding agent can author prompts, scorers, and classifiers, configure the Topics pipeline, build monitor views, create alerts, and run evals, all without leaving your agent environment. Read more →

Your coding agent already knows how you work. It's set up with your repo, tools, and skills. Asking you to move part of that workflow into our in-product agent, Loop, is a high bar, so we have to give you something meaningfully better in return. To keep ourselves honest, we ran both options through a sample of 5 production investigations. Loop was 39% faster, but the insight quality was nearly identical. You should make your decision based on where you want to work next. Use Loop when you want fast, cited (down to the span) production investigations, and you're already in the UI. Especially useful for PMs and subject matter experts. Use the MCP + your coding agent when you want to take your insights into your repo to make changes. Read the full report →

Your team deserves more from your agent observability platform - a single, connected place for instrumentation, investigation, and measurement, enhanced with intelligence. In Braintrust, you can start with an open-ended prompt about agent behavior and carry the investigation into code, evals, automation, and monitoring. New tools like Patterns and Debugger make it easier to find recurring behavior, get to root causes, and act on what you learn. The context you build in one step automatically follows to the next. Work in the platform with an enhanced Loop experience, or move to your coding agent of choice. Read more →

Use the Braintrust JavaScript/TypeScript SDK for tracing and evaling agents in any JS or TS project. Includes integrations for OpenAI Agents, OpenTelemetry, and Temporal. Run evals with a single CLI command or add automatic instrumentation with no code changes. Read more →

Cloudflare Agents now emits native OpenTelemetry traces, and you can route them straight to Braintrust. Export spans via OTLP, instrument your Workers in JavaScript using the Cloudflare Agents SDK, @cloudflare/ai-chat, @cloudflare/think, or Flue. Turn production traces into eval datasets, compare models or prompts, and check whether a change improved the agent before you deploy. Read more →

Imagine doing your job without ever looking anything up on the internet. That's an agent without web search. Giving agents the web changed what they can do, especially on anything recent that isn't reflected in training data. But agents don't search like people, so optimizing their performance is a new challenge. There's a lot of great research out there on search behavior, but we wanted to answer a more operational question: when should search be on, and how should you configure it? We evaluated 1,329 current events questions across 4 models and 14 conditions, comparing @youdotcom, provider built-in search, and no search. We found that: - Search reduced the gap between models from 47.9 points to 5.6 - Retrieval gain declined with event age, from ~45 points for recent events to ~24 for the oldest - Runs with 5+ searches scored 19–49%. A fifth query was associated with lower performance Read the research →

The Braintrust eval library has a repo of skills your coding agent can read, so you can easily build and run evals on your data. Here's an example of using a skill in Claude Code to compare the Codex CLI and Pi, both running GPT-5.6 Sol, on a 30-task stratified SWE-bench Verified dataset. The skill saves those experiments and traces to Braintrust. Try it yourself → See more skills →

Topics clusters every production trace to surface what agents are doing. But 100% coverage only works with a small model that clears the quality bar. So we worked with @baseten to build a benchmark from real Topics traces, iterated against the failure modes that mattered most, and found the model that met our needs. Anyone running a model over production data for their own product can do the same. Read more →

We built a Braintrust-native eval in collaboration with Baseten to test whether GLM-5.2 can preserve exact long-context retrieval under production serving constraints. GLM-5.2's retrieval score is effectively flat as context grows from 25K to 50K, which is the result users most want to see from a sparse-attention long-context model. Read the full GLM-5.2 eval →