Laminar
Open source observability for AI agents
About Laminar
Laminar is an observability platform for AI agents. It captures LLM calls, tool calls, sub-agents, token counts and costs, then renders them as a readable transcript of what the agent actually did rather than a wall of spans you have to reconstruct by hand.
The feature that separates it from ordinary tracing is Signals, which analyses every run looking for failure modes nobody defined in advance. That addresses the real problem with agent debugging: with a normal service you know what failure looks like and write an alert for it, whereas an agent fails in ways you did not anticipate, by looping, by calling the wrong tool, by confidently producing something plausible and wrong. You cannot write assertions for failures you have not imagined yet, so something has to go looking. Laminar is open source, which matters for a tool that necessarily sees your prompts and outputs, since self-hosting is the answer when sending that data to a vendor is not acceptable. The free tier is genuinely usable for a side project at 1 GB and 7-day retention, though the single seat and short retention make it a prototyping tier rather than a production one.
You instrument your agent with the SDK, and Laminar records each run as a trace: the model calls, the tools invoked, any sub-agents spawned, tokens consumed and cost accrued. The trace view presents that as a transcript you can read top to bottom, which is the point, because the useful question is usually "what did it decide to do" rather than "what was the p99 latency". Signals then examines runs and surfaces patterns that look like failures, so problems find you rather than waiting for you to write the right query. Retention and data volume vary by tier, and Signals usage draws on a credit allowance included in each plan.
- •Readable Transcript View - LLM calls, tool calls, sub-agents, tokens and costs presented as a narrative rather than raw spans
- •Signals - Automatic analysis that surfaces failure modes you did not define in advance
- •Open Source - Self-hostable, which is the answer when prompts and outputs cannot leave your infrastructure
- •Sub-Agent Tracing - Nested agent calls are captured as structure, not flattened into noise
- •Cost and Token Tracking - Spend attributable to specific runs rather than a single monthly surprise
- •Usable Free Tier - 1 GB and 7-day retention, enough to instrument a real side project
Anyone running agents in production who has tried to debug one from application logs and given up. It is most valuable when agents are long-running, call several tools, or spawn sub-agents, since those are the cases where reconstructing behaviour by hand becomes hopeless. Teams handling sensitive prompts should look at self-hosting rather than the cloud tiers. If your product is a single model call with no tools, this is more machinery than the problem justifies and simple request logging will do.
Pricing
$0 - $150/mo
- Free$0/mo
- Starter$30/mo
- Pro$150/mo
- EnterpriseContact sales
From the vendor pricing page, 2026-09-21
