Sequential Orchestration: The Pipeline Pattern

Lesson 5 — Multi-agent Pattern 1: chaining agents in a fixed linear order

The Pattern

Sequential orchestration chains agents in a predefined, linear order. Each agent processes the output of the previous one. You — the developer — define the order. The agents don't decide the sequence.

Source: Microsoft Azure — AI Agent Orchestration Patterns

When to Use Sequential

Real-World Example: Contract Generation Pipeline

Microsoft's architecture guide describes a law firm using sequential orchestration for contract generation — 4 agents in a pipeline:

Build It: BA Document Pipeline

Here's a 3-agent sequential pipeline for BA work — draft a BRD, review it, then polish it:

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

llm = ChatOpenAI(model="deepseek-chat", api_key=os.getenv("DEEPSEEK_API_KEY"),
                 base_url="https://api.deepseek.com", temperature=0)

# Agent 1: Draft BRD from requirements
draft_prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a Business Analyst. Draft a BRD from the given requirements. "
               "Include: overview, scope, user stories, acceptance criteria."),
    ("user", "{requirements}"),
])

# Agent 2: Review the draft for completeness
review_prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a senior BA reviewer. Review the BRD for: missing user stories, "
               "vague acceptance criteria, unclear scope. List issues found."),
    ("user", "{draft}"),
])

# Agent 3: Polish the BRD based on review feedback
polish_prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a BA editor. Fix all issues identified in the review. "
               "Output the final polished BRD in markdown."),
    ("user", "Original draft:\n{draft}\n\nReview issues:\n{review}"),
])

# --- The Pipeline (sequential, fixed order) ---
def run_pipeline(requirements: str) -> str:
    # Step 1: Draft
    print("Step 1: Drafting BRD...")
    draft = (draft_prompt | llm).invoke({"requirements": requirements}).content
    
    # Step 2: Review
    print("Step 2: Reviewing draft...")
    review = (review_prompt | llm).invoke({"draft": draft}).content
    
    # Step 3: Polish
    print("Step 3: Polishing BRD...")
    final = (polish_prompt | llm).invoke({"draft": draft, "review": review}).content
    
    return final

result = run_pipeline("Build a login page with OTP, forgot password, and social login")

Notice: you define the order (draft → review → polish). The agents can't deviate. This is a workflow, not an agent — but it's a multi-agent workflow.

Practice Drill

1. Copy the code above into ba-work-agent/sequential_pipeline.py
2. Add your DeepSeek API key
3. Run it with different requirements inputs
4. Try adding a 4th agent: "Format as PDF-ready markdown"
5. Question: Could you do this with a SINGLE LLM call? What would you lose? (Answer: you'd lose the independent review perspective — the same LLM reviewing its own work is less effective)