What are LangChain Deep Agents? | Demystifying AI in 5 Minutes

// TL;DR

Deep Agents = LangGraph + planning + file system + sub-agents + memory. One create_deep_agent() call. Production-ready out of the box.

# Shallow vs Deep

Shallow Agent

✕ Forgets after 5 steps

✕ No planning capability

✕ Context window fills up

✕ Single-threaded execution

✕ No cross-session memory

LLM + tools in a loop

Deep Agent

✓ Coherent over 200+ steps

✓ Built-in todo planning

✓ Virtual file system offloads context

✓ Parallel sub-agent delegation

✓ Persistent memory across sessions

Same loop + 4 superpowers

Claude Code, Manus, and Deep Research all use this pattern. Deep Agents packages it into a library.

# The 3-Layer Stack

Not a replacement for LangGraph. A layer on top. All three compose.

Deep Agents

Full harness. Planning + FS + sub-agents + memory. Start here.

create_deep_agent()

create_agent

Minimal ReAct loop. Model + tools = working agent.

create_react_agent()

LangGraph

Raw graph runtime. Full control over everything.

StateGraph()

# The 4 Superpowers

Auto-injected middleware. You get all four with zero config.

Planning

write_todos tool

Agent creates and checks off tasks. Stays focused across 50+ steps. Same pattern as Claude Code.

Virtual File System

read/write with permissions

Offloads large results. Stores artifacts. Context window never blows up.

Sub-Agents

isolated context workers

Delegate tasks to parallel workers. They run isolated, return results. Parent context stays clean.

Long-term Memory

LangGraph Store

Remembers preferences and decisions across sessions. Gets better with every use.

# .md Files Are the Brain

The real config layer isn't Python or YAML. It's plain markdown.

CLAUDE.md

AGENTS.md

rules/*.md

skills/*.md

📝

System Prompt Core instructions

⚡

Skills On-demand capabilities

📚

Rules Guardrails & constraints

🧠

Memory Learned preferences

📋

Plans Task strategies

🎯

Context Domain knowledge

Change the markdown, change the agent. No redeployment. No code changes. Just edit a file.

Key insight: If you can write a README, you can program a production agent. The .md files are the agent.

# 5 Lines to Start

Everything else is handled by the harness.

main.py

from deepagents import create_deep_agent
from langchain_openai import ChatOpenAI

agent = create_deep_agent(
    model=ChatOpenAI(model="gpt-4o"),
    tools=[web_search, code_executor],
    system_prompt="You are a research expert.",
)

# Planning, file system, sub-agents, memory - all built in.
result = await agent.ainvoke({
    "messages": [{"role": "user", "content": "Research quantum computing breakthroughs in 2026"}]
})

# Pick Your Layer

Start at the top. Drop down only when needed.

You need...	Use this
Quick chatbot, single tool	`create_agent`
Research, code gen, multi-step work	`Deep Agents`
Custom control flow, complex branching	`LangGraph`
Both: harness + custom sub-graph	`Compose them`

// Bottom Line

Deep Agents is what happens when you package the Claude Code architecture into a library. Planning, file system, sub-agents, memory. One pip install. If you're building anything beyond a wrapper, this is your starting point in 2026.

NEXT EPISODE

Thursday

#02 Research Paper

Is Grep All You Need? Agent Harness Paper

A research team tested grep vs vector search inside AI agent loops. The harness matters more than retrieval.

Enjoyed this?

New episodes Mon, Thu, Sun.

All Episodes Follow