LangChain Agents — Playground

01

A chain follows a script. An agent decides.

A chain runs the steps you wrote, in the order you wrote them. An agent loop runs until the model stops asking for things. Same question, same model — press play and count the steps.

Try it

"Where is my order 4471, and can I still return it?"

A chain

from langchain_core.output_parsers import StrOutputParser

chain = prompt | model | StrOutputParser()
chain.invoke({"question": q})

0press play

An agent

from langchain.agents import create_agent

agent = create_agent(model=M, tools=T, system_prompt=S)
agent.invoke({"messages": [{"role": "user", "content": q}]})

0press play

Same model, same question. One of them has somewhere to go and look.

02

Three arguments, one agent.

Everything above is built by a single function. Click each argument in create_agent to see what it carries and what breaks without it — then flip the switch on system_prompt and meet the error every pre-1.0 tutorial will hand you.

Click the code

What Python says when you run it

03

The docstring IS the description.

The model never reads your function body. It reads the tool description — which @tool lifts straight out of your docstring. Pick one of the three and watch the model decide.

Try it

What the model is handed at every turn

The model's decision log

tool calls — wasted calls — answer —

Your docstring is a prompt. It's read on every single turn, by a model choosing between three tools — and it is the only thing standing between a clean call and a guess.

04

What the agent actually passes around.

In and out, an agent's state is one key: {"messages": [...]}. Step through the eight messages that one run of our question produced.

Step through it

result["messages"]

05

Memory is a thread_id.

An agent forgets everything between calls unless you hand it a checkpointer. After that, a thread_id decides which conversation you're continuing — or whether you're a stranger.

Try it

Turn 1 — already sent, thread customer-4471

humanWhere is order 4471?

ai · after 1 tool callOrder 4471 — wireless earbuds, ₹2,499 — was delivered on 12 Sept.

4 messages saved to the checkpoint

Turn 2 — the follow-up

humanCan I still return it?

messages restored — resolved "it" —

Old tutorials will tell you to attach ConversationBufferMemory. It's deprecated and lives in langchain-classic now. Memory stopped being an object you hand the chain and became state the runtime saves, keyed by a thread.

06

The five things tutorials still get wrong.

LangChain 1.0 landed in October 2025 and most blog posts predate it. Five APIs you'll still be handed, and what replaced each. Most of the old ones survive in langchain-classic — while LCEL is the one thing here that never went anywhere.

Check your code

One rule for reading any LangChain tutorial: if it imports from langchain.chains or langchain.memory, it predates 1.0. The modern surface is small — create_agent, @tool, messages, a checkpointer.