langgraph-fundamentals
GitHubLangGraph开发指南，涵盖StateGraph、节点、边及状态管理。用于构建复杂Agent工作流，提供设计方法论与使用场景对比。
Trigger Scenarios
Install
npx skills add langchain-ai/langchain-skills --skill langgraph-fundamentals -g -y
SKILL.md
Frontmatter
{
"name": "langgraph-fundamentals",
"description": "INVOKE THIS SKILL when writing ANY LangGraph code. Covers StateGraph, state schemas, nodes, edges, Command, Send, invoke, streaming, and error handling."
}
- StateGraph: Main class for building stateful graphs
- Nodes: Functions that perform work and update state
- Edges: Define execution order (static or conditional)
- START/END: Special nodes marking entry and exit points
- State with Reducers: Control how state updates are merged
Graphs must be compile()d before execution.
Designing a LangGraph application
Follow these 5 steps when building a new graph:
- Map out discrete steps — sketch a flowchart of your workflow. Each step becomes a node.
- Identify what each step does — categorize nodes: LLM step, data step, action step, or user input step. For each, determine static context (prompt), dynamic context (from state), retry strategy, and desired outcome.
- Design your state — state is shared memory for all nodes. Store raw data, format prompts on-demand inside nodes.
- Build your nodes — implement each step as a function that takes state and returns partial updates.
- Wire it together — connect nodes with edges, add conditional routing, compile with a checkpointer if needed.
| Use LangGraph When | Use Alternatives When |
|---|---|
| Need fine-grained control over agent orchestration | Quick prototyping → LangChain agents |
| Building complex workflows with branching/loops | Simple stateless workflows → LangChain direct |
| Require human-in-the-loop, persistence | Batteries-included features → Deep Agents |
State Management
| Need | Solution | Example |
|---|---|---|
| Overwrite value | No reducer (default) | Simple fields like counters |
| Append to list | Reducer (operator.add / concat) | Message history, logs |
| Custom logic | Custom reducer function | Complex merging |
Nodes
Node functions return partial state updates. Signatures for configuration and runtime access differ by language; use the applicable implementation reference.
Edges
| Need | Edge Type | When to Use |
|---|---|---|
| Always go to same node | add_edge() |
Fixed, deterministic flow |
| Route based on state | add_conditional_edges() |
Dynamic branching |
| Update state AND route | Command |
Combine logic in single node |
| Fan-out to multiple nodes | Send |
Parallel processing with dynamic inputs |
Command
Command combines state updates and routing in a single return value. Fields:
update: State updates to apply (like returning a dict from a node)goto: Node name(s) to navigate to nextresume: Value to resume afterinterrupt()— see human-in-the-loop skill
Python: Use Command[Literal["node_a", "node_b"]] as the return type annotation to declare valid goto destinations.
TypeScript: Pass { ends: ["node_a", "node_b"] } as the third argument to addNode to declare valid goto destinations.
Warning: Command only adds dynamic edges — static edges defined with add_edge / addEdge still execute. If node_a returns Command(goto="node_c") and you also have graph.add_edge("node_a", "node_b"), both node_b and node_c will run.
Send API
Fan-out with Send: return [Send("worker", {...})] from a conditional edge to spawn parallel workers. Requires a reducer on the results field.
Running Graphs: Invoke and Stream
Call graph.invoke(input, config) to run a graph to completion and return the final state.
| Mode | What it Streams | Use Case |
|---|---|---|
values |
Full state after each step | Monitor complete state |
updates |
State deltas | Track incremental updates |
messages |
LLM tokens + metadata | Chat UIs |
custom |
User-defined data | Progress indicators |
Error Handling
Match the error type to the right handler:
| Error Type | Who Fixes | Strategy | Example |
|---|---|---|---|
| Transient (network, rate limits) | System | RetryPolicy(max_attempts=3) |
add_node(..., retry_policy=...) |
| LLM-recoverable (tool failures) | LLM | ToolNode(tools, handle_tool_errors=True) |
Error returned as ToolMessage |
| User-fixable (missing info) | Human | interrupt({"message": ...}) |
Collect missing data (see HITL skill) |
| Unexpected | Developer | Let bubble up | raise |
Core boundaries
- Return partial state updates from nodes instead of mutating state directly.
- Route loops through a named node;
STARTis entry-only. - Define reducers for accumulated list fields; otherwise, the last write wins.
- Account for static edges when using
Commandwithgoto, because both routes execute.
Implementation references
If writing, modifying, or debugging LangGraph code, determine the project's language from its existing files, then read the applicable reference before implementing:
- For Python, read references/python.md.
- For TypeScript, read references/typescript.md.
Read both only when the task covers both languages. For conceptual questions that require no code, do not load either reference.
Version History
-
88df7d9
Current 2026-09-22 01:52
将LangGraph示例拆分为参考文档
- f3ea282 2026-08-02 21:50
