Have any questions?
Building an Agentic
SEO Workflow:
A Python Tutorial
BenStop Prompting. Start Engineering.
In 2023, the cutting edge of SEO was “Prompt Engineering”—learning the magic words to get ChatGPT to write a decent blog post.
In 2026, Prompt Engineering is a commodity skill. The new frontier is Agentic SEO Workflow Engineering.
A single LLM session is linear and limited. It hallucinates, forgets context, and struggles with complex multi-step tasks.
An Agentic SEO Workflow is a system of specialized AI agents chaining tasks together autonomously.
- Agent A (The Researcher) browses the live web to find facts.
- Agent B (The Drafter) writes the content based only on those facts.
- Agent C (The Editor) critiques the draft against specific brand guidelines.
At kōdōkalabs, we do not manually write content. We architect these agentic chains. This approach allows us to scale “Information Gain” without sacrificing quality.
This guide is not a theory piece. It is a technical tutorial. We will walk you through the logic (and the Python code) required to build your first 3-Node SEO Agent Chain using LangChain.
Part 1: Why Agentic SEO Workflows Beat "Chat" SEO
1. The "Context Window" Discipline
When you ask one model to “Research X and then Write Y,” you pollute its context window. It tries to hold the research data and the writing style instructions simultaneously, often degrading performance on both.
In an Agentic chain, the Researcher passes only the clean, verified facts to the Drafter. The Drafter never sees the messy search results, only the structured insight. This “Separation of Concerns” leads to higher quality output.
2. Hallucination Guardrails
A standalone model loves to please you, even if it has to invent facts.
By inserting an Editor Agent whose only job is to verify claims against the input data, you create an adversarial check that catches hallucinations before a human ever sees the draft.
3. Loop Logic (The "Self-Correction" Mechanism)
If the Editor Agent rejects the draft, we can program a “Loop” that sends it back to the Drafter with specific feedback (“Too passive, rewrite section 2”). The system iterates until the quality threshold is met.
Part 2: The Stack
To build this, you need a modern SEO Ops stack. We will use:
- Python 3.10+: The language of AI engineering.
- LangChain: The orchestration framework to connect LLMs.
- OpenAI API (GPT-4o): The reasoning engine.
- Tavily API or Serper.dev: The search tool (to give the Researcher live web access).
Prerequisites: Basic knowledge of Python and API key management.
Part 3: The Architecture
We are building a Sequential Chain:
[User Topic] -> [Researcher Agent] -> {Research Brief} -> [Drafter Agent] -> {First Draft} -> [Editor Agent] -> {Final Manuscript}
Step 1: Setting Up the Environment
First, install the necessary libraries.
pip install langchain langchain-openai tavily-python
import os
from langchain_openai import ChatOpenAI
from langchain_community.tools.tavily_search import TavilySearchResults
# Configuration
os.environ["OPENAI_API_KEY"] = "sk-..." # Your OpenAI Key
os.environ["TAVILY_API_KEY"] = "tvly-..." # Your Tavily Search Key
# Initialize the LLM (Reasoning Engine)
llm = ChatOpenAI(model="gpt-4o", temperature=0)
# Initialize the Search Tool
search_tool = TavilySearchResults(max_results=5) Step 2: Building the Researcher Agent
The goal of this agent is not to write. Its goal is to find “Information Gain”—statistics, quotes, and recent data points.
from langchain.prompts import ChatPromptTemplate
from langchain.schema.output_parser import StrOutputParser
# Define the Researcher Persona
researcher_prompt = ChatPromptTemplate.from_template(
"""
You are a Senior SEO Researcher. Your goal is to find unique, high-value data points on the topic: {topic}.
Do NOT write a blog post.
Do NOT summarize generic knowledge.
1. Use the search tool to find recent statistics, expert quotes, and contrary opinions from 2024-2025.
2. Compile a "Research Brief" containing the top 5 distinct facts with their source URLs.
3. Focus on data that provides "Information Gain" (stuff not found in a generic AI answer).
Output Format:
- Fact 1: [Detail] (Source: URL)
- Fact 2: [Detail] (Source: URL)
...
"""
)
# In a real LangChain agent, we would bind the search tool here.
# For this tutorial simplicity, assume the 'search_results' are passed in via a tool call loop.
# This chain takes the raw search results and synthesizes the Brief.
researcher_chain = researcher_prompt | llm | StrOutputParser() Step 3: Building the Drafter Agent
drafter_prompt = ChatPromptTemplate.from_template(
"""
You are a Technical Content Writer.
Task: Write a blog post section based ONLY on the following Research Brief.
Research Brief:
{research_brief}
Guidelines:
1. Use Markdown formatting (H2, H3, bullet points).
2. Adopt a "Bottom Line Up Front" (BLUF) style. Direct answers first.
3. Cite the sources provided in the brief using Markdown links.
4. Do not invent new facts not present in the brief.
5. Maintain a professional, authoritative tone (E-E-A-T).
Write the content now.
"""
)
drafter_chain = drafter_prompt | llm | StrOutputParser() Step 4: Building the Editor Agent (The Critic)
editor_prompt = ChatPromptTemplate.from_template(
"""
You are a Ruthless Editor in Chief.
Review the following draft:
{draft}
Check for:
1. Fluff words ("In today's fast-paced world", "Unlock the potential").
2. Passive voice.
3. Structural issues (Are there clear H2s?).
If the draft is good, output the finalized version.
If it has issues, rewrite the specific sections to be punchier and more direct.
"""
)
editor_chain = editor_prompt | llm | StrOutputParser() Part 4: Orchestrating the Chain
def run_seo_agent_chain(topic):
print(f"--- Starting Agentic SEO Workflow for: {topic} ---")
# 1. Research Phase
print(">> Agent 1: Researching...")
# (Simulating the tool call wrapper for brevity)
raw_search_results = search_tool.invoke(topic)
research_brief = researcher_chain.invoke({"topic": topic, "search_data": raw_search_results})
print(f" Research Brief Generated ({len(research_brief)} chars)")
# 2. Drafting Phase
print(">> Agent 2: Drafting...")
initial_draft = drafter_chain.invoke({"research_brief": research_brief})
print(f" Draft Generated ({len(initial_draft)} chars)")
# 3. Editing Phase
print(">> Agent 3: Editing...")
final_post = editor_chain.invoke({"draft": initial_draft})
print("--- Workflow Complete ---")
return final_post
# Execute
# final_content = run_seo_agent_chain("Programmatic SEO trends 2025")
# print(final_content)
Part 5: From Code to Strategy
Why go through this effort? Why not just use ChatGPT Plus?
Because Scalability requires Predictability.
When you run this script for 50 different keywords, you get 50 drafts that all follow the exact same structural logic, use verified citations, and adhere to your brand’s editorial voice.
Advanced Optimizations (The “kōdōkalabs Secret Sauce”)
Once you master this basic chain, here is how we upgrade it for Enterprise clients:
- The “Interlinking” Agent: We insert a step between Drafting and Editing where an agent queries a Vector Database (Pinecone) of the client’s existing content to find 5 relevant internal links to insert.
- The “Schema” Agent: A final step that takes the approved content and generates valid JSON-LD FAQ schema to append to the footer.
- Human-in-the-Loop (HITL): We use LangSmith to pause the chain after the “Drafting” phase. A human pilot reviews the brief, approves it, and then allows the Editor agent to proceed.
Conclusion: Engineering Your Moat
The barrier to entry for “generating text” is zero.
The barrier to entry for “generating high-authority, fact-checked, structurally perfect content at scale” is high. It requires engineering.
If you are a CMO or Head of SEO, your job is no longer just managing writers. It is managing the architecture of your intelligence pipeline.
This is what we build at kōdōkalabs. We don’t just sell you the fish; we build the autonomous trawler.