What counts as 'processed data' for billing?

Processed data includes all trace data ingested and stored in Braintrust. This covers your prompts, responses, tool calls, and any attached metadata.

How do I track my usage and costs?

The dashboard and billing page show real-time usage against your plan limits for traces, scores, and data storage. An estimator tool is also available.

Can I deploy Braintrust on my own infrastructure?

Yes. Enterprise plans support a hybrid deployment where the Brainstore data plane runs in your VPC, while control plane remains managed.

How does the Loop agent generate better prompts?

Loop uses your existing traces and eval datasets as context. It can propose prompt variations, run experiments automatically, and suggest scorers based on identified patterns.

Directory/Search/AI Observability/Braintrust

Braintrust

Braintrust is an active observability platform for instrumenting, understanding, and improving AI agents. It provides a complete loop for agent quality: trace everything in production, run evaluations with LLMs, code, or humans, and automatically discover patterns and failures to drive continuous improvement. Key capabilities include real-time agent tracing, versioned dataset management, automated scoring, and a built-in AI agent (Loop) to optimize prompts and evals. It targets AI engineers, ML teams, and product managers building production agents, offering a unified platform from development to production monitoring. The platform is framework-agnostic with SDKs for Python, TypeScript, Go, and more. It features a proprietary database (Brainstore) optimized for large, nested agent traces and offers enterprise-grade security with SOC 2, HIPAA, and GDPR compliance.

Visit website

Added on

Sep 21, 2026

Agent ObservabilityEval PlatformPattern DiscoveryLoop AI AgentTrace AnalysisSOC 2 Compliant

Product information

Everything worth knowing about Braintrust

The complete picture — from everyday features to the technical detail builders and IT folks dig for.

Core features

What it actually does

Agent Observability

Inspect every agent trace, prompt, response, and tool call in real time. Search across millions of logs and track latency, cost, and quality.

Evaluation Platform

Define what good looks like. Run experiments against real datasets, compare prompts/models, and score outputs with LLMs, code, or humans.

Pattern Discovery

Understand agent behavior at scale. Use intelligence to identify recurring issues and opportunities in production traces.

Loop Agent

AI that helps improve agents. Generates better prompts, scorers, and datasets automatically based on your optimization goals.

Custom Trace Views

Build annotation interfaces that match your team's workflow. Review different types of conversations with no frontend work required.

Technical capabilities

Brainstore Database

A proprietary database designed for large, nested agent traces, enabling fast full-text search, low-latency writes, and span loading.

Native SDKs & Integrations

SDKs for Python, TypeScript, Go, Ruby, C#, and more. Works with any stack, no framework lock-in.

Must watch videos

12:44

Tutorial12:44

Stats for evals: Crash course

This crash course walks through the statistical concepts we use in Braintrust's independent evals to interpret pass/fail comparisons between model A and model B. We cover point estimates, sampling variability, Wilson intervals, effective sample size, survivorship bias, distributions, and paired comparisons. This video will help you write better evals analysis as well as know what to look for when reading research papers. 0:00: interpreting model-eval scores 0:57: Bernoulli trials & binomial distributions 1:36: Point estimates & sampling variability 3:56: Standard errors & confidence intervals 5:19: Wilson intervals & the rule of three 7:02: Effective sample size & clustered data 8:25: Missing data & survivorship bias 9:33: Distributions, worst runs & tails 10:52: Paired comparisons & McNemar's test 12:18: Takeaways Homepage 👉 https://www.braintrust.dev/home Changelog 👉 https://www.braintrust.dev/docs/changelog GitHub 👉 https://github.com/braintrustdata/ Discord 👉 https://discord.com/invite/braintrust Newsletter 👉 https://www.braintrust.dev/newsletter X/Twitter 👉 https://x.com/braintrust YouTube 👉 https://www.youtube.com/channel/UCkidH5J6yxUyY2w6PevE37Q LinkedIn 👉 https://www.linkedin.com/company/braintrust-data/

Product updates: Patterns, Loop automations and blind reviews

2:08

Review2:08

Product updates: Patterns, Loop automations and blind reviews

Six new Braintrust updates make active observability, human review, and access control easier to manage: 0:00 - Intro 0:12 - Patterns 0:38 - Loop automations 0:55 - Blind human reviews 1:10 - AI provider permissions 1:26 - Group scopes for online scoring 1:46 - Monitoring views → Dashboards 1:59 - Changelog & closing Homepage 👉 https://www.braintrust.dev/home Changelog 👉 https://www.braintrust.dev/docs/changelog GitHub 👉 https://github.com/braintrustdata/ Discord 👉 https://discord.com/invite/braintrust Newsletter 👉 https://www.braintrust.dev/newsletter X/Twitter 👉 https://x.com/braintrust YouTube 👉 https://www.youtube.com/channel/UCkidH5J6yxUyY2w6PevE37Q LinkedIn 👉 https://www.linkedin.com/company/braintrust-data/

Agent observability with Loop, Patterns, and Debugger

4:20

Tutorial4:20

Agent observability with Loop, Patterns, and Debugger

What does active observability look like in practice? In this video, we use Braintrust Loop to investigate recurring agent behavior across 478 Tau2 support traces, turn the evidence into a reusable pattern, and create a classifier to measure it across future traffic. We cover trace-level debugging, natural-language investigation, evidence-backed reports, reusable patterns, classifier results, and Slack automations. 0:00 - Intro: Active observability 0:28 - From one trace to a scalable workflow 0:50 - Exploring the Exgentic trace dataset 1:15 - Identifying an unresolved agent failure 1:38 - Investigating recurring behavior with Loop 2:19 - Reviewing trace evidence in the Debugger 2:32 - Creating an evidence-backed report 2:43 - Saving findings as reusable patterns 3:00 - Building a classifier for the pattern 3:41 - Routing premature stops with automations 3:55 - Active observability recap Homepage 👉 https://www.braintrust.dev/home Changelog 👉 https://www.braintrust.dev/docs/changelog GitHub 👉 https://github.com/braintrustdata/ Discord 👉 https://discord.com/invite/braintrust Newsletter 👉 https://www.braintrust.dev/newsletter X/Twitter 👉 https://x.com/braintrust YouTube 👉 https://www.youtube.com/channel/UCkidH5J6yxUyY2w6PevE37Q LinkedIn 👉 https://www.linkedin.com/company/braintrust-data/

Who uses it

Real use cases, no hype

Production Agent Debugging

Trace and inspect live agent failures, tool calls, and latency issues.

Prompt & Model Evaluation

Run side-by-side experiments with automated (LLM/code) and human scoring.

Failure Pattern Discovery

Automatically surface recurring issues and behaviors across millions of traces.

AI Quality Regression Testing

Build and version eval datasets from real production failures.

Cross-Team AI Monitoring

Shared dashboards and custom views for engineering and product teams.

What's great

Full-stack agent observability (trace, evaluate, discover)
Turn production traces into eval datasets with one click
AI-powered 'Loop' agent for prompt/scorer optimization
Enterprise security: SOC 2, HIPAA, GDPR, RBAC, hybrid deploy

Technical strengths

Built-in Brainstore DB for fast querying of large, nested traces
Native SDKs for major languages; framework-agnostic integration

Where it falls short

Free tier has tight limits (14-day data retention, 1GB storage)
Core platform fee jumps to $249/mo for Pro features
Data overages ($4/GB) can get expensive at scale

Technical limitations

Advanced security (SSO, RBAC) locked behind paid tiers
Brainstore performance claims unverified against alternatives

For technical folks

The deep-dive specs

FAQ

Includes technical Q&A

Starter for trying it out. Pro for AI-native teams needing custom charts and priority support. Enterprise for high-volume or privacy-sensitive deployments with custom RBAC and SLAs.