create_agent, parameter by parameter

create_agent is the constructor for the whole domain. Learn its shape once and three modules get easier.

The full parameter surface

Positional first, then everything else keyword-only:

create_agent(
    model,
    tools,
    *,
    system_prompt,
    middleware,
    response_format,
    state_schema,
    context_schema,
    checkpointer,
    store,
    interrupt_before,
    interrupt_after,
    debug,
    name,
    cache,
    transformers,
)

Look at that list and notice how much of it you already understand. state_schema, checkpointer, store, interrupt_before, interrupt_after: all ground school vocabulary, handed straight through to the runtime underneath. create_agent is not hiding the graph from you, it is giving you a form to fill in.

The genuinely new ones for this course are middleware (module 2 lives there) and context_schema.

That last one is worth thirty seconds because it sits next to a parameter you already know and does something different. state_schema describes graph state: the thing that flows through the run, gets merged by reducers, and changes as nodes execute. context_schema describes runtime context: the configuration handed in for a given invocation, which the agent reads rather than mutates. Two schemas, one mutable and one not, and having both on the same signature is exactly the kind of adjacency an exam builds a question out of.

And it returns a compiled StateGraph, as established. Nothing exotic comes back.

The parameter that does not exist

Deprecated is not removed, and the difference matters

create_react_agent and its TypeScript twin createReactAgent are deprecated and relocated, still exported, still functional, and emitting a deprecation warning. They are not removed. Code calling them today runs today.

The precise picture, which is worth holding exactly:

• Python: langchain.agents exports only AgentState and create_agent. langgraph.prebuilt still exports create_react_agent, and calling it warns at runtime that it has moved.
• TypeScript: createReactAgent is absent from the langchain agents index, but still exported from langgraph-core carrying an @deprecated JSDoc annotation.

Two different mechanisms doing the same job in two languages: a runtime warning on one side, a compile-time editor squiggle on the other.

Why the precision is the point

There is a failure mode where you read "deprecated" and upgrade it in your head to "deleted", then confidently tell a client their working code is broken. Or you go the other way, see it still exported, and keep writing it into new projects.

Both are wrong and the exam can test either direction. Teach and write create_agent. Describe create_react_agent as legacy but functional. That sentence is defensible against anyone who goes and checks, which is the actual bar.

import { createAgent } from "langchain";

const agent = createAgent({
  model,
  tools: [search, summarise],
  systemPrompt: "You are a research assistant.",
  middleware: [],
});

Note the camelCase. Python is system_prompt, TypeScript is systemPrompt, and that same casing split runs through every single name in the next module.