An open-source AI Agent observability and engineering platform providing automated fault clustering, human-AI alignment evaluation, and CI-level simulated testing.
Latitude is an observability and engineering platform designed specifically for AI Agents and Large Language Models. At its core, it features a three-level tracing model (Span → Trace → Session) capable of capturing complete multi-turn conversations and tool-calling paths.
For fault diagnosis, the platform automatically clusters failed traces into trackable Issues without requiring regex or rule configuration, accurately identifying Agent-specific failure modes like hallucinations and context loss. For quality assurance, Latitude provides a standardized 0-1 scoring system that supports auto-generating evaluation scripts from human annotations and continuously tracks the drift between machine evaluations and human judgments. These generated scripts can be directly used in simulated tests within local environments or CI/CD pipelines to prevent regressions from reaching production.
The underlying project utilizes a DDD-layered Monorepo architecture. The data layer combines PostgreSQL (OLTP), ClickHouse (OLAP), and Weaviate (vector retrieval) to handle different dimensional workloads. Integration is flexible, supporting native TypeScript and Python SDKs as well as the standard OpenTelemetry protocol, with deep integrations with mainstream toolchains like OpenAI, Anthropic, Vercel AI SDK, LangChain, and Claude Code.
Note: The current GitHub
mainbranch is Latitude v2 Alpha; APIs and data formats may change. For production, use thelatitude-v1branch.