Cheatsheet

LangGraph agent cheatsheet (2026 production patterns)

Production LangGraph patterns: state machine design, node types, persistent checkpointing, human-in-the-loop, sub-agents, tracing, eval integration, and the antipatterns to skip.

State machine basics

LangGraph models agents as graphs of state transitions.

Item	Description	Example
`State`	TypedDict / Pydantic model holding everything the graph needs to pass between nodes.
`Nodes`	Functions that take state, return state. Can be LLM calls, tool calls, conditionals, or arbitrary Python.
`Edges`	Transitions between nodes. Static, conditional (branching), or interrupt-driven.
`START / END`	Special nodes marking graph entry and termination.

Node types

Item	Description	Example
`LLM call`	Standard inference — chat completion, structured output, function calling.
`Tool call`	Invoke a tool, fold result into state.
`Conditional branch`	Function returning the next node name based on state.
`Sub-graph`	An entire LangGraph as a node — composing larger agents.

Checkpointing

Production graphs persist state across turns.

Item	Description	Example
`MemorySaver`	In-memory checkpointer for dev.
`PostgresSaver / SqliteSaver`	Persistent checkpointer for production. Stores graph state per thread_id.
`Thread IDs`	Each conversation / session has a thread_id — checkpoints map to it.
`Replay`	Restart from any checkpoint — debug or branch from a specific state.

Human-in-the-loop

Item	Description	Example
`interrupt_before / interrupt_after`	Pause execution before / after specific nodes for human approval.
`Resume with state edit`	Human can edit state mid-pause before resuming. Useful for correction loops.
`Time travel`	Roll back to a previous checkpoint and try a different branch.

Sub-agents + handoffs

Item	Description	Example
`create_react_agent`	Ready-made ReAct agent for many use cases. Wrap as a node in a larger graph.
`Supervisor + workers`	Coordinator agent routes to specialist sub-agents based on intent.
`Handoff`	Explicit node that transfers control to another agent's sub-graph.

Observability

Item	Description	Example
`LangSmith`	First-party tracing — every node, every LLM call, every state transition logged.
`OpenTelemetry`	Custom tracing for OTel-native shops.
`Evals`	Pair traces with evals so regressions on a node are caught.

Antipatterns

Item	Description	Example
`Mega node`	A single LLM-call node doing too much. Split into smaller focused nodes.
`No checkpoint`	Production graphs without checkpoints can't recover from failure mid-run.
`No max-step cap`	Loops can run away. Always cap step count.

FAQ

Best LangGraph pattern for production?

Planner-executor split, postgres checkpointer, human-in-the-loop on destructive actions, LangSmith tracing.

How do I prevent runaway loops?

Set max step count, use no-progress detection, kill on token budget exceedance.

Is LangGraph too heavy for simple use cases?

Probably — for single-step LLM calls a plain SDK is fine. LangGraph wins past 3-5 step orchestration.

Last updated: 2026-06-01.