What we're building

The same calculator/word-counter agent from Lesson 6 (and the LangGraph course), handling three simulated requests from two different users, each request tagged and labeled for production monitoring, one of them given user feedback afterward, then confirmed findable via a filtered query.

What this reveals

The LangGraph course's own tracing lesson covered watching a graph run step by step, print() statements, and get_graph(), useful while developing locally, on your machine, one run at a time. Production is a different situation entirely: many concurrent requests, from many different users, that you are not watching live, and need to be able to inspect after the fact, hours or days later. This lesson combines three tools this course already built, config metadata/tags (Lesson 6), per-request feedback (Lesson 17), and filtered queries (Lesson 18), into the shape a real production integration actually takes.

The two pieces of metadata that matter most in practice are exactly the ones used here: user_id (who made this request, so you can look up "everything this user experienced") and environment (so production traffic never gets confused with your own local testing runs).

The code, piece by piece

run_id = str(uuid.uuid4())
result = app.invoke(
    {"messages": [HumanMessage(question)]},
    config={
        "run_name": "production_agent_request",
        "tags": ["langsmith-course", "production"],
        "metadata": {"user_id": user_id, "environment": "production"},
        "run_id": run_id,
    },
)

Everything here is a combination of earlier lessons: tags/metadata from Lesson 6, and a pre-assigned run_id from Lesson 17, passed through LangGraph's config instead of langsmith_extra since this is a LangChain/LangGraph-native call, not a raw @traceable function.

client.create_feedback(run_id=run_ids[0], key="user_thumbs_up", score=1)

Identical to Lesson 17, just attached to an agent's run instead of a single traced function's run, feedback doesn't care what kind of run produced the id.

production_runs = list(client.list_runs(project_name=PROJECT_NAME, filter='has(tags, "production")', limit=10))

The same filtered query as Lesson 18, narrowed to the "production" tag specifically, exactly how you'd separate real usage from development traffic in a shared LangSmith project.

Checkpoint

  • Production observability vs. local debugging: print() and get_graph() help while developing; metadata, tags, and feedback help after the fact, at scale, across users.
  • user_id and environment metadata: the two most practically useful labels for separating and attributing production traffic.
  • The same tools, applied to a real agent: Lessons 6, 17, and 18's techniques compose directly onto any LangGraph application, not just the single-function examples they were introduced with.

If anything here still feels unclear, ask before moving to Lesson 22.