langchain-and-langgraph-agent-orchestration
Guides building LLM applications with LangChain's chain/agent abstractions and, for stateful multi-step agents, LangGraph's graph-based orchestration (nodes, edges, cycles, checkpointed persistence, human-in-the-loop interrupts). Use when a user asks to "build this with LangChain," "use LangGraph for a stateful agent," "add persistence/checkpointing to a LangChain agent," "add a human approval step in a LangGraph graph," "my LangChain agent loses state between turns," or is deciding between LangChain's `AgentExecutor`, a LangGraph graph, and a hand-rolled control loop.
selvarajmurugesan90 · Static checks on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add selvarajmurugesan90/ops-engineering-skills --skill langchain-and-langgraph-agent-orchestrationAbout this skill
name: langchain-and-langgraph-agent-orchestration
description: >
Guides building LLM applications with LangChain's chain/agent abstractions
and, for stateful multi-step agents, LangGraph's graph-based orchestration
(nodes, edges, cycles, checkpointed persistence, human-in-the-loop
interrupts). Use when a user asks to "build this with LangChain," "use
LangGraph for a stateful agent," "add persistence/checkpointing to a
LangChain agent," "add a human approval step in a LangGraph graph," "my
LangChain agent loses state between turns," or is deciding between
LangChain's AgentExecutor, a LangGraph graph, and a hand-rolled control
loop.
license: Apache-2.0
compatibility: "Claude Code, GitHub Copilot, OpenAI Codex, Cursor, Gemini CLI"
metadata:
domain: ai-agent
maturity: stable
LangChain and LangGraph Agent Orchestration
Purpose
LangChain provides two things that are easy to conflate: a set of composable
building blocks (prompt templates, model wrappers, retrievers, output
parsers, chained via LangChain Expression Language / Runnable), and a
higher-level AgentExecutor that wraps those blocks into a single-loop,
tool-calling agent. LangGraph is a separate, lower-level runtime built by the
same project specifically for agents whose control flow is not a single
linear loop — it models the agent as an explicit graph of nodes and edges,
supports cycles (a node can route back to an earlier node), and adds two
capabilities AgentExecutor does not have out of the box: durable
checkpointed state (so a run can pause, crash, and resume from its last
checkpoint) and first-class human-in-the-loop interrupts (a graph can pause
at a named node until a human approves, edits, or rejects the pending
state). This skill covers choosing between plain LangChain composition,
AgentExecutor, and LangGraph, and operating LangGraph's persistence and
interrupt features correctly. It is a framework-specific complement to
agent-architecture-design, which
covers the underlying control-flow patterns (ReAct loop, plan-and-execute,
finite-state/graph) in a vendor-neutral way — LangGraph is one concrete
runtime that implements the finite-state/graph pattern described there. For
tool access, LangChain/LangGraph agents can call tools defined directly in
Python or exposed via an MCP server; see
mcp-server-development for building
the tool-serving side, which this skill treats as an external dependency
rather than repeating.
When to use
- Deciding whether a task needs plain LangChain chain composition (a fixed
pipeline, no branching),
AgentExecutor(a single ReAct-style tool-calling loop), or a LangGraph graph (multi-step, branching, cyclical, or needing persistence/human approval). - Building an agent whose steps depend on prior results in ways a single linear chain can't express — retries, conditional branches, or loops back to an earlier step.
- Adding durable state to a LangChain/LangGraph agent so a long-running or multi-session workflow survives a process restart or crash mid-run.
- Adding a human-in-the-loop approval checkpoint before an irreversible tool call in an existing LangGraph graph.
- An agent built with
AgentExecutorloses context, re-does completed work, or can't be paused/resumed, and the team is evaluating migrating it to LangGraph. - Debugging a LangGraph graph that loops indefinitely, gets stuck at an interrupt, or fails to restore state correctly from a checkpoint.
Prerequisites & environment
- Python (LangChain/LangGraph's primary, most mature ecosystem) or
JavaScript/TypeScript (
langchain/langgraphnpm packages, closely mirroring the Python API but with some feature lag) — pick one and check current package versions before starting, since both projects have had breaking changes across major versions (notably the LangChain 0.1 → 0.2/0.3 restructuring that split core, community, and partner packages). - An LLM provider integration package (e.g.
langchain-anthropic,langchain-openai) installed separately fromlangchain-core— recent LangChain versions moved provider-specific code out of the core package. - For LangGraph persistence: a checkpointer backend — the in-memory
MemorySaverfor local development/testing only (state is lost on process exit), or a durable checkpointer (SQLite, Postgres, or a managed backend) for anything that must survive a restart. - Tool definitions the agent will call, either as plain Python functions
decorated with LangChain's
@tool, or proxied from an MCP server via an MCP-to-LangChain adapter — confirm which integration path your LangChain version currently supports before assuming API shape. - Clarity on which parts of the workflow are genuinely cyclical/branching (justifying LangGraph) versus a fixed sequence (better served by a plain LCEL chain) — reach for LangGraph only once a chain's limitations are concrete, mirroring the "justify the split" discipline in agent-architecture-design.
Step-by-step guidance
-
Start with the simplest abstraction that fits the task's shape. A fixed sequence (retrieve → prompt → parse) is a plain LCEL chain:
from langchain_core.prompts import ChatPromptTemplate from langchain_core.output_parsers import StrOutputParser from langchain_anthropic import ChatAnthropic prompt = ChatPromptTemplate.from_messages([ ("system", "Summarize the following ticket in one sentence."), ("human", "{ticket_text}"), ]) model = ChatAnthropic(model="claude-sonnet-4-5", temperature=0) chain = prompt | model | StrOutputParser() result = chain.invoke({"ticket_text": ticket_body})No loop, no tool calls, no branching — a chain is the right tool and adding
AgentExecutoror LangGraph here is unjustified complexity. -
**Use
AgentExecutoronly for a si