Observability built for agents, not just chat completions

What it does

Laminar is an open-source observability platform built specifically for AI agents. It auto-instruments popular frameworks—Vercel AI SDK, LangChain, Browser Use, and others—through an OpenTelemetry-native SDK, then feeds traces into a Rust-backed engine with a built-in SQL editor, dashboards, and dataset tools. You can run it fully self-hosted with Docker or lean on the managed cloud offering.

The interesting bit

The Signals feature lets you define monitoring events with plain natural language, so you can flag logical errors or weird agent behavior by description rather than regex. That, plus an unopinionated eval SDK meant for CI/CD, suggests the team actually ships agents and knows where existing tools fall short.

Key highlights

• OpenTelemetry-native tracing for Vercel AI SDK, Browser Use, Stagehand, LangChain, OpenAI, Anthropic, Gemini, and others with minimal code changes.
• Unopinionated eval SDK and CLI for local or CI/CD runs, with a UI for comparing results.
• Natural-language event definitions for AI monitoring (Signals) to catch custom behavior or logical errors.
• Built-in SQL editor and dashboard builder over traces, metrics, and events; queryable via API.
• Rust-backed realtime engine with ultra-fast full-text search over span data and a gRPC exporter.

Caveats

• The simple Docker Compose setup is explicitly labeled for quickstarts and lightweight usage; production self-hosting requires the fuller compose file or the managed platform.
• AI monitoring (Signals) and frontend AI features require configuring a separate LLM provider in self-hosted mode; they are not zero-config.
• The README steers users toward the managed platform for both quickstarts and production, suggesting self-hosting is viable but not the happy path.

Verdict

If you’re building non-trivial AI agents and need to trace multi-step reasoning, run evals in CI, or self-host your telemetry, Laminar is worth a look. If you just want a quick drop-in logger for LLM token usage, it’s probably overkill.