Common Patterns

• https://www.anthropic.com/research/building-effective-agents

1. Workflow:
   1. Create a scaffolding of predefined code paths around LLM calls
   2. LLMs directs control flow through predefined code paths
2. Agent: Remove this scaffolding (LLM directs its own actions, responds to feedback)

Why Frameworks?

• Implementing these patterns does not require a framework like LangGraph.
• LangGraph aims to minimize overhead of implementing these patterns.
• LangGraph provides supporting infrastructure underneath *any**workflow / agent:
  • Persistence
    • Memory
    • Human-In-The-Loop
  • Streaming
    • From any LLM call or step in workflow / agent
  • Deployment
    • Testing, debugging, and deploying

Augmented LLM

# LLM
from langchain_anthropic import ChatAnthropic
llm = ChatAnthropic(model="claude-3-5-sonnet-latest")

# Schema for structured output
from pydantic import BaseModel, Field
class SearchQuery(BaseModel):
    search_query: str = Field(None, description="Query that is optimized web search.")
    justification: str = Field(
        None, justification="Why this query is relevant to the user's request."
    )

# Augment the LLM with schema for structured output
structured_llm = llm.with_structured_output(SearchQuery)

# Invoke the augmented LLM
output = structured_llm.invoke("How does Calcium CT score relate to high cholesterol?")
print(output.search_query)
print(output.justification)

# Define a tool
def multiply(a: int, b: int) -> int:
    return a * b

# Augment the LLM with tools
llm_with_tools = llm.bind_tools([multiply])

# Invoke the LLM with input that triggers the tool call
msg = llm_with_tools.invoke("What is 2 times 3?")

# Get the tool call
msg.tool_calls

Prompt Chaining

Each LLM call processes the output of the previous one:

• E.g., when decomposing a task into multiple LLM calls has benefit.

Example:

• Take a topic, LLM makes a joke, check the joke, improve it twice