LangGraph vs LangSmith for AI agents: Which Should You Use?

LangGraph and LangSmith solve different problems, and mixing them up leads to bad architecture decisions.

LangGraph is the orchestration layer for building agent workflows with state, control flow, retries, and human-in-the-loop steps. LangSmith is the observability and evaluation layer for tracing, debugging, datasets, and regression testing. If you are building AI agents, use LangGraph for execution and LangSmith for visibility.

Quick Comparison

When LangGraph Wins

• •

  You need real control flow.

  If your agent must branch on tool output, retry failed steps, pause for approval, or loop until a condition is met, LangGraph is the right tool. StateGraph gives you explicit nodes and edges instead of burying logic inside prompt spaghetti.

• •

  You are building a production agent with memory and checkpoints.

  LangGraph’s checkpointing pattern matters when an agent runs across multiple turns or long tasks. With MemorySaver or a custom checkpointer, you can persist state between invocations instead of hoping the model remembers what happened.

• •

  You need deterministic orchestration around tools.

  For bank or insurance workflows like claims triage or KYC review, ToolNode plus structured state beats ad hoc function-calling loops every time. You can define exactly when tools run and what happens when they fail.

• •

  You want human-in-the-loop gates.

  Approval steps are first-class in LangGraph design. If a claims decision or payment action needs review before execution, graph-based interruption is cleaner than bolting review logic onto an agent framework after the fact.

When LangSmith Wins

• •

  You already have an agent and need to see what it is doing.

  LangSmith gives you traces across model calls, tool calls, latency spikes, token usage, and errors. If your current problem is “why did this agent do that?”, start here.

• •

  You need evaluation before shipping changes.

  LangSmith’s datasets and experiments are built for regression testing prompts and agents against known cases. That matters when a small prompt change breaks policy extraction or tool selection.

• •

  Your team needs production monitoring.

  Tracing in LangSmith makes it obvious where failures happen: bad retrieval results, malformed tool inputs, slow LLM calls, or broken chains of thought hidden behind too many abstractions.

• •

  You are iterating on prompts more than orchestration.

  If the core problem is model quality rather than workflow design—say summarization accuracy or extraction consistency—LangSmith is the better investment. It helps you compare runs instead of guessing from user complaints.

For AI agents Specifically

Use LangGraph as the runtime and LangSmith as the control tower. Build the agent workflow in StateGraph, wire tools with ToolNode, persist state with checkpointing, then send everything through LangSmith traces so you can debug and evaluate it properly.

If you have to choose one first: choose LangGraph if you are building the agent itself; choose LangSmith if the agent already exists and you need to make it reliable enough for production. For serious AI agents in regulated environments like banking and insurance, the correct answer is usually both—but never as substitutes for each other.

Keep learning

• •The complete AI Agents Roadmap — my full 8-step breakdown
• •Free: The AI Agent Starter Kit — PDF checklist + starter code
• •Work with me — I build AI for banks and insurance companies

By Cyprian Aarons, AI Consultant at Topiax.