Skip to content

LangGraph and deepagents

LanggraphSourceAdapter reads any LangGraph run — plain StateGraphs, compiled subgraphs, deep agents with their subagents — through LangChain's callbacks: inline, on the caller's thread, no queue, no thread hop. The graph is never wrapped or rebuilt; it only gets a config with one more callback. Install e-track[langgraph].

from etrack.providers.adapters.source.langgraph import LanggraphSourceAdapter

source = LanggraphSourceAdapter(mcp=frozenset({"onto_query", "web_search"}))
tracker = Tracker(EventChannel(bus), source)

with tracker.track(thread="t1", config={"configurable": {"thread_id": "t1"}}) as run:
    await agent.ainvoke({"messages": [("user", "…")]}, run.config)
kwarg type default
mcp frozenset[str] ∅ tool names served over MCP: their spans are mcp, not tool
subagents frozenset[str] {"task"} tools that delegate to a subagent — deepagents' task; the span is named by its subagent_type

What becomes what

LangGraph / LangChain sees Event kind
the root chain run
a chain named after its own langgraph_node node — a compiled subgraph mounted as a node is one span, not two
a chat model or LLM llm, with deltas typed by LangChain's v1 content blocks (text, reasoning, tool-call arguments, media)
a tool · a tool in mcp · the task tool tool · mcp · subagent
a retriever retrieval
a GraphInterrupt one interrupt instant, under the span that raised it, carrying what is asked
dispatch_custom_event(name, data) from a node or a tool a custom instant under the span that dispatched it
sequences, middleware wrappers, other plumbing folded away: a frame's parent is always a span that was emitted

Model calls carry model (provider, name, temperature, max tokens) and, on end, usage (input, output, reasoning, cache read and write tokens) as LangChain reports them.

One id for the run

The source pins the run's id as LangChain's root run id. LangSmith, Langfuse, Phoenix, your logs and the viewer all name the run the same — search one id anywhere.

The run's own facts

Put a run's facts under app. in the config's metadata — agent version, tenant, experiment arm — and they land on the root span's attributes, so every backend shows them:

config = {"metadata": {"app.tenant": "acme", "app.agent_version": "2026.11.3"}}
with tracker.track(thread="t1", config=config) as run:
    ...

The root run event then carries attributes={'app.tenant': 'acme', 'app.agent_version': '2026.11.3'}; every key under app. is lifted, nothing else of the metadata is.

Callbacks you already have

Your config's callbacks — a list or a callback manager, LangSmith's tracer, Langfuse's handler — are kept; e-track's handler is added beside them. A callbacks value of any other shape raises SourceError.

The declared graph

The same compiled graph gives the viewer its map → Graphs & labels:

# nodes and arcs as declared, deep agent nodes included
topology = LanggraphTopologyAdapter().read(agent)