Loading...
Loading...
Your agent answered the user. Or it did the wrong thing, called the wrong tool, hallucinated, or burned half your budget. Without traces, you cannot tell. This course wires MLflow tracing into a real LangGraph agent so every step is queryable, cost-attributed, and replayable.
Engineers are learning here from
Add real observability to a LangGraph agent with MLflow. Trace every node and tool call, attach cost and token metadata, and answer "what did the agent actually do?" without grepping logs. A guided curriculum that takes you from print statements to queryable traces per request.
Stop guessing what your agent did. Trace every step with the open-source observability tool ML teams already use.
What you'll ship
What you'll learn
Curriculum
Foundations of LLM observability
Why observability is different from logging, what a trace is, and where MLflow fits.
The @trace decorator
Build a reusable decorator that wraps any function in an MLflow span with cost and token metadata.
Trace a multi-route agent
Wire MLflow into a LangGraph state machine that routes between SQL, RAG, and web search.
Operational dashboards
Turn the trace data into the dashboards an on-call engineer would actually open.
Who it's for
Your agent works in a notebook, but you do not yet have a way to inspect what it did when something looks wrong.
The data scientists handed over a notebook. You need to instrument it before you can put it behind an API.
You want a working MLflow setup before deciding between MLflow, Langfuse, and Phoenix for your stack.
FAQ
MLflow is the open-source observability tool ML teams already know. This course is the free, broadest entry into LLM tracing. Phoenix is OpenTelemetry-native and focused on agent and RAG observability. Langfuse adds production tracing plus the eval and prompt-management surfaces. Pick MLflow first if your team is already on it. Pick Phoenix or Langfuse for production agent stacks.
No. The source repo ships with a small SQLite database and a sample PDF. The labs use those throughout, and the patterns transfer to any LangGraph agent.
Yes. MLflow tracing is the same tool used to track ML experiments at Databricks scale. The local setup in this course works for development. For production, point it at a managed MLflow tracking server.
The router and tools call any LangChain-compatible chat model. The MLflow @trace decorator and span attributes are model-agnostic. Swap the client; the tracing pattern is identical.
Trace once, debug forever. The mental model travels to any observability stack.