# Laminar — Open-Source Observability for AI Agents

> Open-source observability for AI agents: self-host with Docker Compose, then use the SDK to trace runs, metrics, and outputs end-to-end.

## Install

Save as a script file and run:

## Quick Use

1. Self-host locally:
   ```bash
   git clone https://github.com/lmnr-ai/lmnr
   cd lmnr
   docker compose up -d
   ```
2. Install the Python SDK:
   ```bash
   pip install --upgrade 'lmnr[all]'
   ```

## Intro

Open-source observability for AI agents: self-host with Docker Compose, then use the SDK to trace runs, metrics, and outputs end-to-end.

- **Best for:** agent teams who need repeatable tracing and debugging, not just ad-hoc logs, across prompts, tools, and multi-step runs
- **Works with:** Docker Compose for self-hosting; SDK instrumentation in your agent app
- **Setup time:** 30–60 minutes

## Practical Notes

- Quant: define 3 core metrics per agent workflow (latency, tool-call count, success rate) and baseline them before you optimize prompts.
- Quant: keep a replay set of 20 representative runs; compare traces after every change to detect regressions.

---

## Observability-first iteration

If you can’t answer these questions with data, you’re guessing:

- Which step dominates latency?
- Which tool calls fail most often?
- Which prompt change improved success rate vs just “felt better”?

## Minimal instrumentation strategy

1. Trace every run with a stable run id.
2. Attach tool-call spans with inputs/outputs (redact secrets).
3. Capture final outcomes (pass/fail + reason).

## Don’t drown in dashboards

Start with one workflow and one team. Once the metrics are trusted, scale to more services.

### FAQ

**Q: Do I need to self-host?**
A: No. The repo documents self-hosting; teams can choose managed options or local-only usage.

**Q: What should I instrument first?**
A: One end-to-end workflow that currently fails or is slow—make it measurable.

**Q: How do I compare prompt changes?**
A: Use a fixed replay set and compare traces/metrics, not anecdotes.

## Source & Thanks

> Source: https://github.com/lmnr-ai/lmnr
> License: Apache-2.0
> GitHub stars: 2,875 · forks: 195

---

<!-- ZH -->

## 快速使用

1. 本地自托管：
   ```bash
   git clone https://github.com/lmnr-ai/lmnr
   cd lmnr
   docker compose up -d
   ```
2. 安装 Python SDK：
   ```bash
   pip install --upgrade 'lmnr[all]'
   ```

## 简介

AI Agent 的开源可观测性平台：用 Docker Compose 自托管后，通过 SDK 采集 trace/指标/输出结果，支持端到端追踪、调试与对比评估，让 agent 的迭代从“感觉”变成可量化。

- **适合谁：** 希望把 agent 的追踪与调试做成可复用能力，而不只靠零散日志的团队
- **可搭配：** Docker Compose 自托管；在你的 agent 应用里接入 SDK
- **准备时间：** 30–60 分钟

## 实战建议

- 量化建议：为每条工作流定义 3 个核心指标（延迟、tool 调用次数、成功率），先基线再优化 prompt。
- 量化建议：固定 20 条可复现的回放样本；每次变更后对比 traces，及时发现回归。

## 用可观测性驱动迭代

如果下面问题回答不出数据，你就在“凭感觉优化”：

- 哪一步最耗时？
- 哪个 tool 调用最容易失败？
- 哪次 prompt 变更真的提升成功率？

## 最小接入策略

1. 每次 run 都用稳定的 run id 做追踪。
2. tool 调用做 span 记录输入/输出（注意脱敏）。
3. 记录最终结果（成功/失败 + 原因）。

## 避免看板淹没

先从一个工作流 + 一个团队开始；指标可信后再扩到更多服务。

### FAQ

**必须自托管吗？**
答：不必须。仓库提供了自托管路径；团队也可以选择托管或仅本地使用。

**优先接入什么？**
答：先把一条当前最慢或最不稳定的工作流做成可量化的端到端追踪。

**如何对比 prompt 变更？**
答：用固定回放集对比 traces/指标，而不是靠主观感受。

## 来源与感谢

> Source: https://github.com/lmnr-ai/lmnr
> License: Apache-2.0
> GitHub stars: 2,875 · forks: 195


---
Source: https://tokrepo.com/en/workflows/laminar-open-source-observability-for-ai-agents
Author: Script Depot