📚 AI 工程与智能体教程站

English

简体中文

English

简体中文

Chapter 9 · Deep Agents Introduction: create_deep_agent ​

Chapter goals:

  • Understand Deep Agents positioning: Agent Harness above LangGraph
  • Master basic usage of create_deep_agent()
  • Learn to configure multiple Providers (OpenAI, Anthropic, Google)
  • Understand basic concepts of Skills and Memory

9.1 What is Deep Agents? ​

Deep Agents is an Agent Harness provided by LangChain, offering on top of LangGraph:

  • Simplified API (create Agent in one line of code)
  • Built-in context management
  • Filesystem tools
  • Sub-agent support
  • Memory system

Comparison: LangGraph is Runtime, Deep Agents is Harness (Harness layer)

9.2 Quick Start ​

python
from deepagents import create_deep_agent

# Create agent with one line
agent = create_deep_agent(
    model="gpt-4o",
    tools=["search", "code", "filesystem"],
    memory=True
)

# Run agent
result = agent.run("Research and summarize this topic")

9.2.1 Provider Configuration ​

python
from deepagents import create_deep_agent
from deepagents.providers import OpenAIProvider, AnthropicProvider, GoogleProvider

# Multiple providers
agent = create_deep_agent(
    providers=[
        OpenAIProvider(api_key="sk-..."),
        AnthropicProvider(api_key="sk-ant-..."),
        GoogleProvider(api_key="AIza..."),
    ],
    default_provider="openai"
)

9.3 Skills System ​

Skills are reusable Agent capabilities and instructions:

markdown
<!-- .deepagents/skills/researcher/AGENTS.md -->
---
name: researcher
description: Research topics and generate reports
model: anthropic:claude-haiku-4-5-20251001
---

You are a research assistant. Your tasks:
1. Search for relevant information
2. Synthesize findings
3. Generate structured report

9.3.1 Loading Skills ​

python
from deepagents import load_skills

skills = load_skills(".deepagents/skills/")
agent = create_deep_agent(skills=skills)

9.4 Memory System ​

9.4.1 Short-term Memory ​

Conversation history within a session:

python
agent = create_deep_agent(
    memory="short_term",  # Last N messages
    memory_window=10
)

9.4.2 Long-term Memory ​

Persistent memory across sessions:

python
agent = create_deep_agent(
    memory="long_term",
    memory_store="sqlite:///memory.db"
)

9.5 Approval Modes ​

Three approval modes for human-in-the-loop:

ModeBehavior
ManualEvery tool call requires approval
AutoAuto-approve safe operations
YOLONo approval, maximum autonomy
python
agent = create_deep_agent(
    approval_mode="manual",  # or "auto" or "yolo"
    allowed_tools=["read_file", "search"]
)

Chapter Summary ​

  1. Deep Agents: High-level Harness built on LangGraph;
  2. create_deep_agent(): One-line Agent creation with configurable providers;
  3. Skills: Reusable capabilities defined in AGENTS.md;
  4. Memory: Short-term (session) and long-term (persistent) options;
  5. Approval Modes: Manual, Auto, YOLO for human oversight control.

🛠️ Hands-on Practice ​

  1. Create a Deep Agent with multiple provider fallback (OpenAI → Anthropic → Google).

  2. Define custom Skills in AGENTS.md format and load them into an Agent.

  3. Implement human approval workflow with Manual mode for sensitive operations.

🧪 随堂测验

点击你认为正确的选项。答错时会展示正确答案与原因解析。

1. What is Deep Agents in relation to LangGraph?

2. How do you create a Deep Agent?

3. What are Skills in Deep Agents?

4. Which approval mode gives maximum Agent autonomy?