Langchain

LangChain spent two years as both the most-used and most-criticized framework in AI — then 1.0 (October 2025) did the unusual thing: it agreed with the critics and shrank. The repo now calls itself "the agent engineering platform," and the framework's center is one well-built thing instead of forty abstractions.

Highlights

• create_agent — a standard, production-grade tool-calling agent loop, running on the LangGraph runtime, in one call.
• Middleware — hooks at every step of the loop, with built-ins that matter: human-in-the-loop approval, context summarization, PII redaction.
• Normalized content blocks — reasoning traces, citations, and tool calls in one shape across providers; the swap-the-model promise made real.
• In-loop structured output — typed results generated inside the agent loop, not via an extra LLM call.
• The ecosystem moat — the largest integration surface in the category (langchain-* packages for every model, vector store, and tool), in Python and JS.
• A clean escalation path — drop down to LangGraph for custom graphs and durable state; out to LangSmith for tracing and evals.

In an AI-assisted workflow

pip install langchain        # or: npm install langchain
# agent = create_agent(model, tools, middleware=[HumanInTheLoop()])

The 2026 fit: teams that want a standard agent loop without marrying a provider, and that value the graduated stack (LangChain → LangGraph → LangSmith) over assembling equivalents.

Good to know

MIT, ~139k stars, free; the company monetizes LangSmith (freemium per-seat). Where it sits against the data-framework lineage of LlamaIndex — the classic confusion — is exactly the LangChain vs LlamaIndex question; the wider field is Agent Frameworks in 2026.