Laminar is an open-source AI agent observability platform focused on tracing, evaluating, and debugging AI agents. It is designed to help developers discover and fix agent failures at scale.
Key capabilities:
- Automatic capture: Laminar automatically captures LLM calls, tool calls, sub-agents, costs, and tokens, and presents them in a readable transcript view.
- Failure discovery: Laminar Signals analyzes every agent run to reveal unexpected failure patterns. Similar failures are clustered into distinct patterns, providing a high-level overview. Signal clusters show the full distribution of agent failures and behaviors, helping determine whether a newly reported issue has occurred before.
- Cost-effectiveness: flow-1 is a fine-tuned and RL post-trained model for intelligent and efficient trace analysis, with trace analysis intelligence comparable to GPT-6 Sol, while analyzing 20 times more traces per dollar.
- Slack integration: Laminar notifies about new failures and clusters via Slack, and you can mention Laminar in Slack to ask any question about traces.
- Coding agent support: Laminar CLI and MCP provide coding agents with full SQL access to Laminar, enabling them to investigate traces, build evaluations, and verify fixes.
- Regression detection: Each fixed error cluster can be automatically converted into an evaluation dataset. Run evaluations after changes to catch regressions, so you can iterate with confidence.
- Easy integration: Integrate with Claude Agent SDK, OpenAI Agents SDK, LangChain DeepAgents, and others in two lines of code.
- Automatic input extraction: Laminar extracts the agent task from every run, making it easy to create end-to-end evaluation datasets from production traces.
- Custom dashboards: Build dashboards with custom SQL queries to track statistics from traces and signals.
- Full SQL access: Query all platform data with raw SQL. Coding agents can query data via MCP or CLI.
- Data annotation interface: Build labeled datasets from traces, dataset rows, or SQL results for evaluation and fine-tuning.
- Browser agent screen recording: Capture the agent's browser session and display it alongside the trace.
- Full-text search: Full-text search across every span's inputs, outputs, and attributes.
- Storage optimization: Laminar stores only unique content from agent runs, enabling faster ingestion and 20x more convenientConvenient storage.
- Open source and deployment: open source from day one, Apache 2.0 license. Can be set up via Docker, deployed on AWS or GCP using Helm charts, or run on Kubernetes using the Laminar Helm chart.
- Enterprise-ready: HIPAA and SOC 2 Type II compliant. Redacts sensitive information from every span (PII redaction) before storing it on Laminar's own infrastructure.
How to use:
- Integration: integrate with supported frameworks in two lines of code.
- Deployment: set up with Docker, or deploy on AWS, GCP, or Kubernetes using Helm charts.
- Interaction: receive notifications and query via Slack, or use the CLI and MCP to give coding agents SQL access.
- Evaluation: build evaluation datasets from traces, and run evaluations after changes to detect regressions.
Use cases:
- Teams running AI agents at scale that need to monitor and debug agent behavior.
- Scenarios that need to discover unexpected failure patterns and perform cluster analysis.
- Developers who need to build evaluation datasets from production traces and perform regression testing.
- Organizations that need enterprise-grade compliance (HIPAA, SOC 2 Type II) and PII redaction.
- Agent development using frameworks such as Claude Agent SDK, OpenAI Agents SDK, and LangChain DeepAgents.
Limitations:
- No specific pricing information is provided in the material.
- No other potential limitations or system requirements are provided in the material.
- No supported programming languages or specific environment details are provided in the material.
Laminar is led by CEO Magnus Müller, and its customers say they run millions of agent sessions in the cloud, and when problems arise, Laminar's trace view is the first place they look.
