Langgraph Code

LangGraph

Source: https://langchain-ai.github.io/langgraph/llms.txt (this harness's index into the official docs — fetch and re-verify against it if this content seems stale; LangGraph moves fast).

When to Reach for LangGraph

LangGraph models an agent as an explicit state machine — nodes and edges over a typed state object — rather than a linear chain of prompts. Reach for it when the task needs: multi-step control flow with branching, durable state that survives a crash or process restart, or a point where a human must approve/edit before the agent continues. For a simpler single-agent tool-calling loop with no explicit state machine, langchain-code's create_agent is the right level of abstraction instead — read that skill first if you're unsure LangGraph's added control is actually needed.

Core Workflow

1. Define state as a TypedDict (or dataclass/Pydantic model) — see references/graph-api.md.
2. Add nodes (plain functions that take state, return partial updates) and edges (static or conditional) — see references/graph-api.md.
3. Compile the graph — must call .compile() before it's runnable.
4. If the graph needs to survive a restart, resume mid-conversation, or pause for human approval: compile with a checkpointer and always pass a thread_id — see references/persistence-and-checkpointing.md. This is opt-in, not automatic — a graph compiled without a checkpointer has no persistence at all.

Gotchas (do not skip)

• Checkpointing is not automatic. You must explicitly pass a checkpointer to .compile() and a thread_id in every invocation's config. A graph with neither has zero crash-resume capability.
• InMemorySaver does not persist across process restarts — development-only. Use PostgresSaver (or another durable backend) for anything that must survive a real crash.
• Node re-execution on resume: when resuming from a checkpoint, LangGraph re-runs the interrupted node from its start, not from where it left off mid-function. Design node bodies to be idempotent (safe to re-run) — do not put a side effect like "send an email" directly in a node body without a guard.
• Don't mix static and conditional edges from the same node. If a node has both a plain add_edge and a add_conditional_edges targeting it, both paths fire — pick one routing style per node.
• State updates overwrite by default. If a state field should accumulate (e.g. a running message list) rather than replace, it needs an explicit reducer via Annotated[list[X], operator.add] (or LangGraph's add_messages for chat history) — without one, each node's return value replaces the field wholesale.

Not Yet Covered Here (fetch before relying on these topics)

This skill's reference files cover graph construction and persistence/checkpointing in depth. The following official pages exist in the llms.txt index but have not yet been fetched into a reference file — fetch the relevant one before advising on these topics, don't guess: streaming (.../langgraph/streaming), subgraph composition (.../langgraph/use-subgraphs), observability/tracing (.../langgraph/observability), common error troubleshooting (.../langgraph/common-errors), and the dedicated workflows-vs-agents framing (.../langgraph/workflows-agents).