LangChain review

What it is

LangChain is the default framework for building applications on language models. It provides abstractions for prompts, retrieval, memory, tools and agents, with integrations for essentially every model provider and vector database, plus LangGraph for structured agent workflows and LangSmith for tracing.

Where it shines

Ecosystem breadth. Whatever combination of model, vector store, document loader and tool you need, an integration probably exists and someone has written about it. For teams moving fast, that removes weeks of plumbing.

LangGraph is the significant improvement of the past two years. Defining an agent as an explicit graph of nodes and edges — with state you can inspect, checkpoints you can resume from, and branches you can reason about — turned agent development from vibes into engineering. Complex flows with human approval steps, retries and parallel branches are genuinely manageable.

LangSmith is the other reason teams stay. Seeing every prompt, response, token count and latency for a production trace, then building evaluation datasets from real traffic, is how you stop guessing why an agent misbehaved.

Where it falls short

Abstraction cost. LangChain hides what is actually sent to the model, and when output is wrong the layers between your code and the API become an obstacle. Experienced teams often end up reading library source to understand their own prompts.

Version churn is the long-standing complaint. Breaking changes and reorganised packages have burned projects more than once, and tutorials go stale quickly.

There is also a culture of over-adoption: plenty of systems import a large framework to make one API call, and pay for it in complexity and dependency risk.

Pricing in plain English

The libraries are free and MIT licensed. LangSmith Plus is $39 per user per month for tracing and evaluation. Enterprise pricing covers self-hosted LangSmith with SSO and support.

Who should use it

Engineering teams building non-trivial LLM applications — retrieval systems, multi-step agents, anything needing observability. For a simple prompt-and-response feature, call the API directly.