LangChain Framework¶

Scope checked: 2026-09-03. LangChain is now centred on create_agent: a configurable agent harness composed from a model, tools, a system prompt, and middleware. It is useful when an application needs that composition and provider portability; it is not required for a one-call model integration. LangChain overview

Choose the Right Layer¶

Framework choice does not replace authorization, validation, or domain logic. Keep money movement, publication, access control, and irreversible effects in deterministic application code.

Minimal Agent Contract¶

An agent is a model calling tools in a loop. A production harness needs an explicit contract for each layer:

Do not treat a provider-neutral interface as proof that all providers offer identical context windows, tool semantics, structured-output guarantees, or regional availability.

Current Starting Point¶

# pip install -U langchain "langchain[anthropic]"
from langchain.agents import create_agent
from langchain.tools import tool


@tool
def lookup_project(project_id: str) -> dict:
    """Return a public project record after server-side authorization."""
    return {"project_id": project_id, "status": "published"}


agent = create_agent(
    model="anthropic:claude-sonnet-4-6",
    tools=[lookup_project],
    system_prompt="Use tools only for public project records.",
)

result = agent.invoke(
    {"messages": [{"role": "user", "content": "Show project happyin-space"}]}
)

The example intentionally keeps the tool small. The function must still authenticate the caller, validate the project_id, enforce scope, and return a structured error when the action is unavailable. The model should not receive a broad database credential merely because the tool is convenient.

Middleware Is a Policy Boundary¶

Middleware can add routing, retry, tool policy, context management, and guardrails. Treat each addition as executable policy rather than decoration:

1. state the trigger and expected result;
2. name the owner of the policy and the version;
3. make the allow, deny, retry, and escalation paths observable;
4. test the policy with malformed input and unavailable dependencies;
5. keep a deterministic fallback only when it preserves the same safety guarantees.

For example, a retry wrapper may handle a transient model error. It must not silently replay a tool call that could publish, charge, delete, or send a message.

State and Durable Work¶

LangChain agents are built on LangGraph, which provides durable execution, persistence, and human-in-the-loop support. That is useful for long-running work, but persistence changes the data contract: record what state is retained, who can restore it, and how an interrupted external effect is reconciled. LangChain overview

Use a task record outside the model context:

{
  "task_id": "research-044",
  "state": "awaiting_review",
  "input_revision": "source-set@2026-09-03",
  "allowed_actions": ["draft_article"],
  "approval_ref": null,
  "terminal_receipt": null
}

The runtime may summarize this record for the model, but the application remains the authority for transitions, approval, and the terminal receipt.

Observability and Evaluation¶

Tracing can show prompts, tool calls, state transitions, latency, and errors. It does not prove that the output is correct. Pair traces with:

• a representative, versioned evaluation set;
• expected tool-call and refusal cases;
• quality measures tied to the task rather than token count;
• redaction and retention rules for prompts, tool inputs, and outputs;
• a review path for high-impact failures.

LangSmith documents tracing and evaluation for agents built with LangChain, LangGraph, and other frameworks. LangSmith observability

Gotchas¶

• Old examples use deprecated chain and memory helpers. They can still appear in search results. Fix: start from the current create_agent and provider-integration documentation, then pin the package versions you test.
• A common model interface is not a common capability contract. Providers differ in tool calling, limits, and error behavior. Fix: validate the selected model-provider pair in CI or a controlled canary.
• Middleware can hide an effectful retry. A model retry is not necessarily safe for an external action. Fix: put idempotency and reconciliation at the executor boundary.
• Trace data can become a shadow data store. Debug logs often contain prompts and tool outputs. Fix: classify, redact, restrict, and expire observability data.
• State persistence is not a permission system. Restoring a checkpoint must not restore expired authority. Fix: revalidate capabilities and approvals at effect time.

Sources¶

• LangChain overview
• LangChain agents
• LangChain provider integrations
• LangSmith observability

See Also¶

• langgraph
• agent orchestration
• tool use patterns
• agent safety alignment
• llmops