AI Application Development Mastery Part 11: AI Application Design Patterns

1. Core Patterns

Core patterns are the foundational building blocks. Every AI application uses at least one of these. They are simple to implement, well-understood, and form the basis for more complex patterns.

1.1 Prompt-Response Pattern

The simplest AI pattern: send a prompt to an LLM, receive a response. Despite its simplicity, this pattern powers a surprising number of production applications when combined with well-engineered prompts.

# Pattern: Prompt-Response
# Complexity: Low | Latency: Low | Cost: Low
# pip install langchain-openai

import os
from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate

# Requires OPENAI_API_KEY environment variable
# export OPENAI_API_KEY="sk-..."

# Simple prompt-response with structured output
llm = ChatOpenAI(model="gpt-4", temperature=0)

# Email classifier — a pure prompt-response application
email_classifier = ChatPromptTemplate.from_messages([
    ("system", """You are an email classifier. Classify the
    incoming email into exactly one category:
    - URGENT: Requires immediate action
    - SUPPORT: Customer support request
    - SALES: Sales inquiry or lead
    - INTERNAL: Internal team communication
    - SPAM: Unwanted or irrelevant

    Respond with JSON: {{"category": "...", "confidence": 0.0-1.0,
    "reasoning": "..."}}"""),
    ("human", "{email_text}")
])

chain = email_classifier | llm

# Usage
result = chain.invoke({
    "email_text": "Our production database is down and customers "
                  "cannot access their accounts. Please fix ASAP!"
})
# Output: {"category": "URGENT", "confidence": 0.95,
#          "reasoning": "Production outage affecting customers"}

1.2 RAG (Retrieval-Augmented Generation) Pattern

The most important pattern in AI application development. RAG augments the LLM's knowledge with external data by retrieving relevant documents before generating a response.

# Pattern: RAG (Retrieval-Augmented Generation)
# Complexity: Medium | Latency: Medium | Cost: Medium
# pip install langchain-openai langchain-chroma

import os
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_chroma import Chroma
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnablePassthrough

# Requires OPENAI_API_KEY environment variable
# export OPENAI_API_KEY="sk-..."

# Components
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")
vectorstore = Chroma(
    collection_name="company_docs",
    embedding_function=embeddings,
    persist_directory="./chroma_db"
)
retriever = vectorstore.as_retriever(
    search_type="mmr",  # Maximum Marginal Relevance
    search_kwargs={"k": 5, "fetch_k": 20}
)

llm = ChatOpenAI(model="gpt-4", temperature=0)

# RAG prompt template
rag_prompt = ChatPromptTemplate.from_messages([
    ("system", """Answer the user's question using ONLY the
    provided context. If the answer is not in the context,
    say "I don't have information about that in our docs."

    Context:
    {context}"""),
    ("human", "{question}")
])

# RAG chain
def format_docs(docs):
    return "\n\n---\n\n".join(
        f"Source: {d.metadata.get('source', 'Unknown')}\n"
        f"{d.page_content}" for d in docs
    )

rag_chain = (
    {"context": retriever | format_docs,
     "question": RunnablePassthrough()}
    | rag_prompt
    | llm
)

# Usage
answer = rag_chain.invoke("What is our refund policy?")

1.3 Tool Use Pattern

The LLM decides which external tools to call, constructs the arguments, interprets the results, and formulates a response. This pattern gives LLMs the ability to take actions in the real world.

# Pattern: Tool Use
# Complexity: Medium | Latency: Medium-High | Cost: Medium
# pip install langchain langchain-openai

import os
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool
from langchain.agents import create_tool_calling_agent, AgentExecutor
from langchain_core.prompts import ChatPromptTemplate

# Requires OPENAI_API_KEY environment variable
# export OPENAI_API_KEY="sk-..."

@tool
def get_weather(city: str) -> str:
    """Get the current weather for a city."""
    # In production: call a real weather API
    return f"Weather in {city}: 72F, sunny, humidity 45%"

@tool
def calculate(expression: str) -> str:
    """Evaluate a mathematical expression safely."""
    allowed = set("0123456789+-*/.()")
    if all(c in allowed or c == ' ' for c in expression):
        return str(eval(expression))
    return "Invalid expression"

@tool
def search_database(query: str) -> str:
    """Search the company database for information."""
    # In production: query your actual database
    return f"Database results for '{query}': 42 matching records"

# Create tool-calling agent
llm = ChatOpenAI(model="gpt-4", temperature=0)
tools = [get_weather, calculate, search_database]

prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant with access to tools. "
               "Use them when needed to answer questions accurately."),
    ("human", "{input}"),
    ("placeholder", "{agent_scratchpad}")
])

agent = create_tool_calling_agent(llm, tools, prompt)
executor = AgentExecutor(agent=agent, tools=tools, verbose=True)

# The LLM decides which tools to call
result = executor.invoke({
    "input": "What is the weather in Tokyo and what is 15% of 2500?"
})

2. Intermediate Patterns

Intermediate patterns combine core patterns with state management, persistence, and more complex data flows. These patterns power most production AI applications.

2.1 Chat + Memory Pattern

Maintains conversation history across turns, enabling contextual multi-turn conversations. The LLM references previous messages to provide coherent, context-aware responses.

# Pattern: Chat + Memory
# Complexity: Medium | Latency: Low-Medium | Cost: Medium
# pip install langchain-openai

import os
from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.chat_history import InMemoryChatMessageHistory
from langchain_core.runnables.history import RunnableWithMessageHistory

# Requires OPENAI_API_KEY environment variable
# export OPENAI_API_KEY="sk-..."

llm = ChatOpenAI(model="gpt-4", temperature=0.7)

# Prompt with memory placeholder
prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful AI assistant. Use the "
               "conversation history to provide contextual answers."),
    MessagesPlaceholder(variable_name="history"),
    ("human", "{input}")
])

chain = prompt | llm

# Session-based memory store
store = {}

def get_session_history(session_id: str):
    if session_id not in store:
        store[session_id] = InMemoryChatMessageHistory()
    return store[session_id]

# Wrap chain with message history
chat_with_memory = RunnableWithMessageHistory(
    chain,
    get_session_history,
    input_messages_key="input",
    history_messages_key="history"
)

# Multi-turn conversation
config = {"configurable": {"session_id": "user-123"}}
r1 = chat_with_memory.invoke(
    {"input": "My name is Alice and I work at Acme Corp."},
    config=config
)
r2 = chat_with_memory.invoke(
    {"input": "What company do I work at?"},  # Uses memory
    config=config
)
# r2 correctly answers "Acme Corp" from memory

2.2 Document QA Pattern

A specialized combination of RAG and Chat + Memory optimized for question-answering over a specific document corpus. Includes source citation, confidence scoring, and follow-up question handling.

# Pattern: Document QA
# Complexity: Medium | Latency: Medium | Cost: Medium
# Combines: RAG + Chat + Memory + Source Citation
# pip install langchain-openai langchain-chroma pydantic

import os
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_chroma import Chroma
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import JsonOutputParser
from pydantic import BaseModel, Field

# Requires OPENAI_API_KEY environment variable
# export OPENAI_API_KEY="sk-..."

class QAResponse(BaseModel):
    answer: str = Field(description="The answer to the question")
    sources: list[str] = Field(description="Source documents used")
    confidence: float = Field(description="Confidence score 0-1")
    follow_up: list[str] = Field(
        description="Suggested follow-up questions"
    )

llm = ChatOpenAI(model="gpt-4", temperature=0)
parser = JsonOutputParser(pydantic_object=QAResponse)

doc_qa_prompt = ChatPromptTemplate.from_messages([
    ("system", """You are a document QA assistant. Answer questions
    using ONLY the provided context documents. Always cite which
    source document(s) you used. Rate your confidence from 0 to 1.
    Suggest 2-3 follow-up questions the user might ask.

    {format_instructions}

    Context Documents:
    {context}"""),
    ("human", "{question}")
])

doc_qa_chain = (
    doc_qa_prompt.partial(
        format_instructions=parser.get_format_instructions()
    )
    | llm
    | parser
)

# Returns structured response with sources, confidence, follow-ups

2.3 Workflow Automation Pattern

Chains multiple AI steps together with conditional logic, data transformations, and external system integrations. This pattern is the backbone of AI-powered business process automation.

# Pattern: Workflow Automation
# Complexity: Medium-High | Latency: High | Cost: Medium-High
# Best implemented with: n8n, Zapier, LangGraph
# pip install langgraph langchain-openai

import os
from typing import TypedDict
from langgraph.graph import StateGraph, END
from langchain_openai import ChatOpenAI

# Requires OPENAI_API_KEY environment variable
# export OPENAI_API_KEY="sk-..."

class WorkflowState(TypedDict):
    email_text: str
    classification: str
    sentiment: str
    response_draft: str
    approved: bool

llm = ChatOpenAI(model="gpt-4", temperature=0)

def classify_email(state: WorkflowState) -> dict:
    """Step 1: Classify the incoming email."""
    response = llm.invoke(
        f"Classify this email as SUPPORT, SALES, or BILLING: "
        f"{state['email_text']}"
    )
    return {"classification": response.content.strip()}

def analyze_sentiment(state: WorkflowState) -> dict:
    """Step 2: Analyze customer sentiment."""
    response = llm.invoke(
        f"Rate sentiment as POSITIVE, NEUTRAL, or NEGATIVE: "
        f"{state['email_text']}"
    )
    return {"sentiment": response.content.strip()}

def draft_response(state: WorkflowState) -> dict:
    """Step 3: Draft an appropriate response."""
    response = llm.invoke(
        f"Draft a {state['sentiment'].lower()} tone response "
        f"for this {state['classification']} email: "
        f"{state['email_text']}"
    )
    return {"response_draft": response.content}

def route_by_sentiment(state: WorkflowState) -> str:
    """Route negative sentiment to human review."""
    if state["sentiment"] == "NEGATIVE":
        return "human_review"
    return "auto_send"

# Build workflow graph
wf = StateGraph(WorkflowState)
wf.add_node("classify", classify_email)
wf.add_node("sentiment", analyze_sentiment)
wf.add_node("draft", draft_response)

wf.set_entry_point("classify")
wf.add_edge("classify", "sentiment")
wf.add_edge("sentiment", "draft")
wf.add_conditional_edges("draft", route_by_sentiment, {
    "human_review": END,  # Human reviews negative responses
    "auto_send": END      # Auto-send positive/neutral
})

workflow = wf.compile()

3. Advanced Patterns

Advanced patterns involve autonomous decision-making, iterative self-improvement, and multi-agent coordination. They are the most powerful but also the most complex and expensive to run.

3.1 Agent Loop Pattern

The agent receives a goal, then enters a think-act-observe loop, iteratively taking actions and refining its approach until the goal is achieved or a termination condition is met.

# Pattern: Agent Loop (ReAct / Think-Act-Observe)
# Complexity: High | Latency: High | Cost: High
# The agent iterates until the goal is achieved
# pip install langgraph langchain-openai

import os
from typing import TypedDict, Annotated
from langgraph.graph import StateGraph, END
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage, SystemMessage
import operator

# Requires OPENAI_API_KEY environment variable
# export OPENAI_API_KEY="sk-..."

class AgentLoopState(TypedDict):
    messages: Annotated[list, operator.add]
    goal: str
    plan: str
    iteration: int
    completed: bool

llm = ChatOpenAI(model="gpt-4", temperature=0)

def think_node(state: AgentLoopState) -> dict:
    """Agent thinks about what to do next."""
    messages = [
        SystemMessage(content=f"""You are working on: {state['goal']}
        Current plan: {state.get('plan', 'None')}
        Iteration: {state['iteration']}

        Analyze the current situation and decide:
        1. What has been accomplished so far?
        2. What remains to be done?
        3. What is the next specific action?
        4. Should we FINISH? (say GOAL_COMPLETE if yes)"""),
        *state["messages"][-5:]
    ]
    response = llm.invoke(messages)
    completed = "GOAL_COMPLETE" in response.content
    return {
        "messages": [response],
        "completed": completed,
        "iteration": state["iteration"] + 1
    }

def act_node(state: AgentLoopState) -> dict:
    """Agent takes an action based on its thinking."""
    last_thought = state["messages"][-1].content
    messages = [
        SystemMessage(content="Execute the next action from "
                      "the plan. Provide concrete output."),
        HumanMessage(content=f"Action to take: {last_thought}")
    ]
    response = llm.invoke(messages)
    return {"messages": [response]}

def should_continue(state: AgentLoopState) -> str:
    if state.get("completed", False):
        return "end"
    if state.get("iteration", 0) >= 10:
        return "end"
    return "act"

# Build agent loop graph
loop = StateGraph(AgentLoopState)
loop.add_node("think", think_node)
loop.add_node("act", act_node)

loop.set_entry_point("think")
loop.add_conditional_edges("think", should_continue, {
    "act": "act",
    "end": END
})
loop.add_edge("act", "think")  # Loop back

agent_loop = loop.compile()

3.2 Planning Pattern

The agent first creates an explicit plan (a sequence of steps), then executes the plan step-by-step, re-planning when necessary. This separates strategic thinking from tactical execution.

# Pattern: Planning (Plan-and-Execute)
# Complexity: High | Latency: High | Cost: High
# Separates planning from execution for better results
# pip install langgraph langchain-openai

import os
from typing import TypedDict
from langgraph.graph import StateGraph, END
from langchain_openai import ChatOpenAI
from langchain_core.output_parsers import JsonOutputParser

# Requires OPENAI_API_KEY environment variable
# export OPENAI_API_KEY="sk-..."

class PlanState(TypedDict):
    goal: str
    plan: list  # List of step descriptions
    current_step: int
    step_results: dict
    needs_replan: bool

llm = ChatOpenAI(model="gpt-4", temperature=0)

def planner_node(state: PlanState) -> dict:
    """Create or revise the plan."""
    if state.get("needs_replan", False):
        prompt = f"""The original plan failed at step
        {state['current_step']}. Results so far:
        {state['step_results']}

        Revise the plan for goal: {state['goal']}
        Return a JSON list of remaining steps."""
    else:
        prompt = f"""Create a step-by-step plan to achieve:
        {state['goal']}

        Return a JSON list of 3-7 specific, actionable steps.
        Example: ["Research topic X", "Write outline", ...]"""

    response = llm.invoke(prompt)
    # Parse JSON list from response
    import json
    try:
        plan = json.loads(response.content)
    except json.JSONDecodeError:
        plan = [response.content]

    return {
        "plan": plan,
        "current_step": 0,
        "needs_replan": False
    }

def executor_node(state: PlanState) -> dict:
    """Execute the current step of the plan."""
    step_idx = state["current_step"]
    step = state["plan"][step_idx]

    response = llm.invoke(
        f"Execute this step thoroughly: {step}\n"
        f"Previous results: {state.get('step_results', {})}"
    )

    results = state.get("step_results", {})
    results[f"step_{step_idx}"] = response.content

    return {
        "step_results": results,
        "current_step": step_idx + 1
    }

def check_progress(state: PlanState) -> str:
    if state["current_step"] >= len(state["plan"]):
        return "done"
    return "execute"

# Build plan-and-execute graph
plan_graph = StateGraph(PlanState)
plan_graph.add_node("planner", planner_node)
plan_graph.add_node("executor", executor_node)

plan_graph.set_entry_point("planner")
plan_graph.add_edge("planner", "executor")
plan_graph.add_conditional_edges("executor", check_progress, {
    "execute": "executor",
    "done": END
})

plan_execute = plan_graph.compile()

3.3 Multi-Agent Orchestration Pattern

Multiple specialized agents collaborate on a task, coordinated by a supervisor or through peer-to-peer communication. This pattern was covered extensively in Part 10, but here we place it in the context of the full pattern catalog.

# Pattern: Multi-Agent Orchestration
# Complexity: Very High | Latency: Very High | Cost: High
# See Part 10 for detailed implementations

# Summary of multi-agent sub-patterns:
MULTI_AGENT_PATTERNS = {
    "supervisor": {
        "description": "Central coordinator routes to specialists",
        "agents": "3-8",
        "best_for": "Quality-gated pipelines, dev workflows",
        "framework": "LangGraph (best), CrewAI, AutoGen"
    },
    "swarm": {
        "description": "Autonomous agents with handoff protocols",
        "agents": "2-6",
        "best_for": "Customer service routing, triage",
        "framework": "LangGraph, OpenAI Swarm"
    },
    "debate": {
        "description": "Adversarial agents improve via argumentation",
        "agents": "2-4",
        "best_for": "Analysis, evaluation, high-stakes decisions",
        "framework": "LangGraph, AutoGen"
    },
    "hierarchical": {
        "description": "Tree of managers delegating to workers",
        "agents": "5-20+",
        "best_for": "Large projects, enterprise automation",
        "framework": "CrewAI (hierarchical process), LangGraph"
    }
}

4. Pattern Selection Guide

Choosing the right pattern and framework is the most impactful architectural decision you will make. This section provides a systematic approach to pattern selection.

4.1 Framework-Pattern Matrix

This matrix shows which framework is best suited for each design pattern:

4.2 Decision Tree

Follow this decision tree to select the right pattern for your use case:

# AI Pattern Selection Decision Tree

START: What does your application need to do?
|
|-- Single LLM call sufficient?
|   |-- YES: Do you need external data?
|   |   |-- NO  --> Prompt-Response (LangChain/direct API)
|   |   |-- YES --> RAG Pattern (LangChain + vector DB)
|   |-- NO: Continue...
|
|-- Does it need conversation history?
|   |-- YES: Is it over a document corpus?
|   |   |-- YES --> Document QA (LangChain)
|   |   |-- NO  --> Chat + Memory (LangChain)
|   |-- NO: Continue...
|
|-- Does it need to take actions (APIs, DB, files)?
|   |-- YES: Is it a single action or a chain?
|   |   |-- Single  --> Tool Use (LangChain)
|   |   |-- Chain   --> Workflow Automation (LangGraph/n8n)
|   |-- NO: Continue...
|
|-- Does it need autonomous multi-step reasoning?
|   |-- YES: Does it need explicit planning?
|   |   |-- YES --> Planning Pattern (LangGraph)
|   |   |-- NO  --> Agent Loop (LangGraph)
|   |-- NO: Continue...
|
|-- Does it need multiple specialized agents?
|   |-- YES --> Multi-Agent Orchestration (LangGraph/CrewAI)
|
|-- Is the user non-technical?
|   |-- YES --> n8n (self-hosted) or Zapier (cloud)
|
|-- None of the above?
    --> Start with Prompt-Response and iterate

Framework Selection Summary

5. Anti-Patterns

Anti-patterns are common mistakes that seem reasonable but lead to poor outcomes. Recognizing them saves time, money, and frustration.

5.1 Common Mistakes

5.2 Pattern Smells

These signs indicate you may be using the wrong pattern:

# Pattern smell detection checklist

PATTERN_SMELLS = {
    "High latency for simple tasks": {
        "likely_cause": "Using Agent Loop for Prompt-Response tasks",
        "fix": "Downgrade to a simpler pattern",
        "metric": "If avg response > 10s for classification tasks"
    },
    "Inconsistent outputs": {
        "likely_cause": "Missing structured output parsing",
        "fix": "Add Pydantic output parser or function calling",
        "metric": "If JSON parse failures > 5% of requests"
    },
    "Context window errors": {
        "likely_cause": "Memory Bloat or God Prompt anti-pattern",
        "fix": "Implement summary memory or break prompt into chains",
        "metric": "If token count regularly exceeds 80% of limit"
    },
    "Agent loops without progress": {
        "likely_cause": "Infinite Agent Loop anti-pattern",
        "fix": "Add progress tracking and termination conditions",
        "metric": "If agent takes > 5 iterations for simple tasks"
    },
    "High cost per query": {
        "likely_cause": "Premature Agent-ification",
        "fix": "Audit each agent step - can it be a simple chain?",
        "metric": "If cost/query > $0.10 for routine operations"
    },
    "Retrieval returns irrelevant docs": {
        "likely_cause": "Poor chunking, wrong embedding model, or "
                        "RAG Everything anti-pattern",
        "fix": "Tune chunk size, use re-ranking, audit what needs RAG",
        "metric": "If retrieval relevance score < 0.7 average"
    }
}

6. Exercises & Self-Assessment

Conclusion & Next Steps

You now have a complete catalog of AI application design patterns and a systematic framework for choosing the right one. Here are the key takeaways from Part 11:

• Core patterns — Prompt-Response, RAG, and Tool Use are the fundamental building blocks that every AI developer must master
• Intermediate patterns — Chat + Memory, Document QA, and Workflow Automation combine core patterns with state management and persistence for production-grade applications
• Advanced patterns — Agent Loops, Planning, and Multi-Agent Orchestration enable autonomous, multi-step reasoning but come with significantly higher cost and complexity
• Framework-pattern matrix — LangChain excels at RAG and chains, LangGraph at complex agents, CrewAI at multi-agent teams, n8n/Zapier at no-code workflow automation
• Start simple — The golden rule is to always start with the simplest pattern that could work, then upgrade only when you hit a specific limitation
• Anti-patterns — The God Prompt, Premature Agent-ification, RAG Everything, Infinite Agent Loop, Memory Bloat, Framework Overload, and Tool Explosion are the most common and costly mistakes