Get started: Engineer

What can I do with this system?

• Trace any Python LLM workflow (plain functions, Microsoft Agent Framework, or LangChain) into a single unified Langfuse trace with one SDK call: semantic-tracing.
• Run an offline evaluation of your workflow against a golden dataset, score it with LLM-judge and deterministic metrics via the shared scoring service, and gate on the result, locally or as a CI check, using semantic-evaluation.
• Grow that golden dataset from real production traces through human review, without writing sync code yourself: the shared review-sync service handles the annotation-queue → dataset pipeline.
• Online evaluation (scoring live production traffic as it happens, rather than replaying a golden dataset offline) is coming soon.

Fastest path to a working example

The Workshop is the concrete, hands-on version of all three of the above, built around a small support-assistant agent. It takes about half a day and produces:

1. A traced workflow run showing up in your own Langfuse project.
2. A human-reviewed trace promoted into a golden dataset.
3. A gated offline-evaluation run with real metric scores.

Work through the modules in order:

• Module 01: Setup : connects to the two deployed shared services and installs one of the two equivalent scaffolds (Microsoft Agent Framework or LangChain)
• Module 02: Tracing and Module 03: Human Review
• Modules 04–07 : run manifest, predictor adapter, metric authoring, and running evaluations

Where to go next

• Deep SDK reference: semantic-tracing docs, semantic-evaluation docs
• Service reference: scoring-service docs, review-sync-service docs
• Using Claude Code or Copilot? See Agent Skills: your coding agent sets up and keeps up tracing and evaluation for you.