LangChain

LangChain is the standard framework for chaining LLM components. In 2025, the focus shifted to LangGraph for building stateful, cyclic agents.

When to Use

• Orchestration: Chaining "Prompt -> LLM -> Parser".
• Agents: Using LangGraph to build agents that can loop, retry, and keep state.
• Integrations: 1000+ connectors for vector DBs, APIs, and tools.

Core Concepts

LangGraph

The successor to AgentExecutor. A graph-based way to define agent flows with cycles (loops).

LCEL (LangChain Expression Language)

The declarative pipe syntax: prompt | llm | output_parser.

LangSmith

Observability platform to trace and debug complex chains.

Best Practices (2025)

Do:

• Use LangGraph: For any non-trivial agent. AgentExecutor is legacy.
• Use LCEL: It enables streaming and async out of the box.
• Trace everything: Connect to LangSmith to see why your agent failed.

Don't:

• Don't over-abstract: If a simple Python function works, don't wrap it in a Chain.

References

• LangChain Documentation