Laminar

Teams running AI agents in production struggle to see what their agents are actually doing across many LLM calls, tool calls, and sub-agents, and often cannot tell why an agent failed until a user reports a bad outcome. Failures like repeated redundant tool calls, ignored errors, or hallucinated results can hide inside long traces that are costly and slow to review manually. Without a way to classify unanticipated failure modes or compare a new issue against past ones, teams end up firefighting agent bugs one at a time rather than systematically. Standard LLM observability tools may also be expensive to run at the scale of millions of agent sessions.

Laminar automatically captures LLM calls, tool calls, sub-agent activity, costs, and tokens into a readable transcript trace view. Its "Signals" feature uses a custom model (flow-1) to analyze every agent run and surface failure modes without requiring teams to predefine them, then clusters similar failures so teams can see the full distribution of issues and check whether a new report has occurred before. Laminar notifies teams of new failures and clusters in Slack, and its CLI and MCP give coding agents direct SQL access to traces so they can investigate issues, build evals, and verify fixes. The platform also supports turning fixed error clusters into eval datasets to catch regressions, offers custom dashboards with full SQL access, a UI for data annotation, browser session recording for browser agents, full-text search across trace data, and claims 20x cheaper storage by storing only unique content in agent runs. It is open source under Apache 2.0, can be self-hosted via Docker or deployed on AWS/GCP with Helm charts, and states PII redaction, HIPAA, and SOC 2 Type II compliance for enterprise use.