LangGraph
Your agent now models LangGraph workflows as directed graphs correctly — StateGraph, state schemas, nodes, edges, Command, Send, invoke, streaming, and error handling. Sourced from LangChain AI.
What it installs
| Artifact | Path (in your project) | Purpose |
|---|---|---|
| Skill | .agents/skills/langgraph-fundamentals/ | Protocol the agent follows on LangGraph work — sourced from langchain-ai/langchain-skills |
| Rule | .claude/rules/langgraph.md | Always-loaded pointer: triggers on any LangGraph import, StateGraph, or graph node/edge work |
How it works
LangGraph's power is also its footgun: graphs compose beautifully when state schemas, node signatures, and control flow (Command, Send) are right — and fail opaquely when they aren't. The skill encodes the canonical patterns: declaring state via TypedDict / Pydantic, adding nodes + conditional edges, using Command for explicit state + routing, using Send for map-reduce, wiring checkpointers for persistence, and streaming events to clients.
The rule loads a pointer into every session, so the agent consults the skill before writing any StateGraph or node function — no explicit invocation needed.
Setup
No env vars or external accounts required for the skill. Provider API keys needed to run the underlying LLMs. Skill fetched from GitHub during harness-kit add via npx skills add.
Pairs well with
langchain—create_agent()runtime for individual nodesredis— checkpointer storage for durable graph statecode-review-gates— enforce state-schema and edge-condition checks on every commit