![LangGraph Example: Email Agent
from langgraph.graph import StateGraph
from langchain_anthropic import ChatAnthropic
# Define state
class AgentState(TypedDict):
messages: list[AnyMessage]
preferences: dict
memory_summary: str
# Initialize LLM
llm = ChatAnthropic(model="claude-3-7-sonnet-latest")
# Nodes
def fetch_context_node(state: AgentState):
state["preferences"] = {"tone": "formal"}
return state
def llm_node(state: AgentState):
response = llm.invoke(state["messages"])
state["messages"].append(response)
return state
# Build graph
builder = StateGraph(AgentState)
builder.add_node("fetch_context", fetch_context_node)
builder.add_node("llm", llm_node)
builder.add_edge("fetch_context", "llm")
graph = builder.compile()
NextGen_Outlier 9](https://image.slidesharecdn.com/context-250708213326-e28271d8/85/Context-Engineering-for-AI-Agents-approaches-memories-pdf-9-320.jpg)
Context Engineering for AI Agents, approaches, memories.pdf
- 1. Context Engineering for AI Agents By - Tamanna NextGen_Outlier 1
- 2. What is Context Engineering? Definition: Designing the information ecosystem for AI agents to perform tasks effectively. Unlike prompt engineering, it’s a systems-level approach. Goal: Provide the right info, in the right format, at the right time. Analogy: Like giving a chef the ingredients and recipe for a perfect dish. NextGen_Outlier 2
- 3. Why is it Hard for AI Agents? Limited Context Window: Finite token space (e.g., 8k, 32k tokens). Dynamic Tasks: Multi-step workflows need different context per step. Memory Management: Agents lack inherent memory, unlike humans. Tool Overload: Selecting relevant tools without confusion. Balancing Relevance: Too much or too little context causes issues. NextGen_Outlier 3
- 4. Types of Context Type Description System Instructions Rules defining behavior (e.g., "Act as a support agent"). User Prompt Immediate task (e.g., "Draft an email"). Conversation History Past messages in the session. Long-Term Memory User preferences or facts (e.g., "Prefers formal tone"). Retrieved Information External data via RAG or APIs. Tool Responses Outputs from tools (e.g., "Free at 3 PM"). NextGen_Outlier 4
- 5. Types of Memory Memory Type Description Semantic General facts/preferences (e.g., "Favorite color: blue"). Episodic Specific past interactions (e.g., "Liked pasta recipe"). Procedural Task rules (e.g., "Check calendar before email"). Example: Storing "User prefers formal tone" as semantic memory. NextGen_Outlier 5
- 6. Context Engineering Strategies Write: Save context outside the window (e.g., scratchpad). Select: Retrieve relevant data using RAG or embeddings. Compress: Summarize history to fit token limits. Isolate: Sandbox irrelevant data to avoid confusion. Example: Summarize chat as "User discussed project X, prefers agile." NextGen_Outlier 6
- 7. What is LangGraph? A low-level framework for stateful AI agents. Key Features: Controllable workflows with nodes and edges. Memory management (short/long-term). Tool integration and human-in-the-loop support. Ideal for context engineering due to fine-grained control. NextGen_Outlier 7
- 8. LangGraph Workflow Nodes: Tasks (e.g., fetch context, call LLM). Edges: Flow between tasks. State: Dynamic context (messages, memory). Example Workflow: User query → Fetch context → Call LLM → Store memory → Respond. NextGen_Outlier 8
- 9. LangGraph Example: Email Agent from langgraph.graph import StateGraph from langchain_anthropic import ChatAnthropic # Define state class AgentState(TypedDict): messages: list[AnyMessage] preferences: dict memory_summary: str # Initialize LLM llm = ChatAnthropic(model="claude-3-7-sonnet-latest") # Nodes def fetch_context_node(state: AgentState): state["preferences"] = {"tone": "formal"} return state def llm_node(state: AgentState): response = llm.invoke(state["messages"]) state["messages"].append(response) return state # Build graph builder = StateGraph(AgentState) builder.add_node("fetch_context", fetch_context_node) builder.add_node("llm", llm_node) builder.add_edge("fetch_context", "llm") graph = builder.compile() NextGen_Outlier 9
- 10. Getting Started with LangGraph Install: pip install langgraph langchain langchain-anthropic Set API Key: export ANTHROPIC_API_KEY="your-api-key" Use LangMem for long-term memory (optional). NextGen_Outlier 10
- 11. Practical Tips Use RAG for efficient memory retrieval. Summarize long histories to save tokens. Namespace memories by user or app. Reflect on feedback to improve procedural memory. Test with small context windows first. NextGen_Outlier 11
- 12. Conclusion Context engineering is key to intelligent AI agents. Manage context with Write, Select, Compress, Isolate. Use LangGraph for controllable workflows. Next Steps: Try LangGraph with your own agent project! NextGen_Outlier 12