ISOM 260 · Suffolk University

Multi-Agent Systems
in Practice

One agent is useful. A team of agents changes everything. Watch three specialists collaborate in real-time — and learn when a team beats a solo performer.

Multi-Agent Pipelines Agent Coordination Handoffs Specialization
↓ Scroll to begin · Press Space to advance
Section 01 — The Org Chart Analogy

Companies Have Departments.
Agent Systems Have Specialists.

One person can't do everything well. Neither can one agent. The same organizational principles that drive company structure apply to AI agent design.

Company
Research Dept
Gathers data & market intel
↓
Analysis Dept
Finds patterns & insights
↓
Editorial Dept
Reviews & publishes reports
=
Agent System
Researcher Agent
Searches & structures facts
↓
Analyst Agent
Identifies patterns & insights
↓
Editor Agent
Fact-checks & scores quality
Key Insight
Companies have departments because one person can't do everything. Agent systems work the same way — division of labor, specialization, and coordination costs apply to AI just like human teams.
Section 02 — Three Architecture Patterns

Pick the Right Architecture

Not all multi-agent systems look the same. Three core patterns cover most real-world use cases.

Pipeline

Researcher ↓ Analyst ↓ Editor
Best for: Sequential workflows where each stage builds on the last. Most common pattern.

Debate

Agent A ↔ Agent B ↓ [Judge]
Best for: Complex decisions needing adversarial review. Think red team / blue team.

Orchestrator

Manager /   |   \ W1   W2   W3
Best for: Dynamic task allocation, unknown scope. The pattern behind Claude Code's team feature.
Today's focus: We're building the Pipeline pattern. Researcher → Analyst → Editor. You designed these architectures in Session 6 — today you code them.
Section 03 — Agent Communication

How Agents Talk to Each Other

Three patterns for inter-agent communication, each with different tradeoffs.

Message Passing
# Agent A output → Agent B input research = researcher("AI trends") # Pass output directly analysis = analyst(research) # Chain continues report = editor(analysis)
Shared Context
# All agents read/write shared workspace = { "facts": [], "insights": [], "report": None } researcher(workspace) analyst(workspace) editor(workspace)
Orchestration
# Manager decides routing def manager(task): plan = decide_routing(task) for step in plan: agent = pick_agent(step) result = agent(step) update_plan(result) return compile()
Pattern Simplicity Flexibility Debuggability
Message Passing High — just function calls Low — fixed order High — clear trace
Shared Context Medium — state mgmt High — any agent reads Low — who wrote what?
Orchestration Low — complex setup High — dynamic routing Medium — manager decides
Section 04 — The Pipeline in Action

Watch Data Transform at Each Stage

Follow a query as it flows through each agent, gaining structure, analysis, and quality at every handoff.

Query
User input
"What are the trends
in enterprise AI
adoption for 2026?"
→
Researcher
Structured facts
facts: [
 "73% plan to adopt..."
 "$4.1B market size..."
 "Top 3 use cases..."
]
sources: 5
→
Analyst
Patterns + insights
insights: [
 "Adoption accelerating
  post-2024 AI boom..."
]
recommendations: 3
→
Editor
Verified report + score
report: "Enterprise AI
 adoption is..."
quality_score: 8.5/10
citations: verified
pipeline.py — What You'll Build
def run_pipeline(query): # Stage 1: Researcher gathers structured facts research = call_agent(RESEARCHER_PROMPT, query) # Stage 2: Analyst finds patterns + insights analysis = call_agent(ANALYST_PROMPT, research) # Stage 3: Editor fact-checks + scores quality final = call_agent(EDITOR_PROMPT, research + analysis) return final # verified report with quality score
See it live → Pipeline Demo
Section 05 — Hands-On Checkpoint

Time to Build

45 minutes in Google Colab. Build a Researcher → Analyst → Editor pipeline from scratch.

Multi-Agent Workshop Notebook
Build a team of specialized AI agents that collaborate on research, analysis, and quality control.
Open in Colab Workshop Timer
Phase 1
15 minutes
Build the Researcher agent — searches for information, returns structured findings with sources and confidence levels.
Phase 2
15 minutes
Build the Analyst agent — receives research, identifies patterns, generates insights and recommendations.
Phase 3
15 minutes
Add the Editor agent + run the full pipeline end-to-end. Fact-check, improve clarity, score quality 1–10.
Section 06 — The Telephone Game

Information Distortion Across Agents

Each agent subtly transforms information. After 3 agents, is the output still accurate? Watch a single fact degrade through the pipeline.

Original
"73% of enterprises plan to adopt AI agents by 2026"
Researcher
"73% of enterprises are adopting AI agents" ← subtle shift: "plan to" removed
Analyst
"Most enterprises are rapidly adopting AI agents" ← generalized: 73% → "most", added "rapidly"
Editor
"Enterprise AI adoption is nearly universal" ← distorted: 73% future plan → "nearly universal"
Generalization
Specific numbers become vague qualifiers: "73%" → "most"
Amplification
Adding intensity that wasn't in the source: → "rapidly"
Omission
Dropping qualifiers: "plan to adopt" → "are adopting"
Fabrication
New claims not in the source: → "nearly universal"
Real Stakes
In a research report, distortion is annoying. In healthcare, legal, or financial AI systems, it's dangerous. Every multi-agent system needs verification gates — points where humans or automated checks confirm accuracy before the pipeline continues.
Safeguards: Citation chains (each agent cites where it got each fact), confidence propagation, and human review gates at critical handoffs. If Agent C gives a wrong answer based on Agent A's research — who's accountable?
Section 07 — When Multi-Agent Is Worth It

Multi-Agent Isn't Always Better

Each additional agent adds latency, cost, and complexity. Use the right tool for the right job.

Use Case Single Agent Multi-Agent Winner
Simple Q&A ✓ Fast, cheap ✗ Overkill Single
Research report ⚠ Shallow ✓ Deep, verified Multi
Code review ⚠ One perspective ✓ Adversarial Multi
Data lookup ✓ Direct ✗ Unnecessary Single
Content creation ⚠ Generic ✓ Specialized Multi
The Complexity Tax
Each agent adds ~2–5 seconds latency and costs 1 additional API call. A 3-agent pipeline takes 3x longer and costs 3x more than a single agent. Only pay this tax when the quality gain justifies it.
"Don't use a team of agents when one agent with good tools will do.
But don't use one agent when the task demands expertise you can't fit in a single prompt."
The Multi-Agent Decision Rule
Section 08 — The Stack So Far

Your Growing Toolkit

Each session adds a new capability. Here's what you can build now.

S9 Your Industry UPCOMING
S8 Multi-Agent Systems ◀ YOU ARE HERE
S7 RAG — Knowledge Agents
S6 ReAct + Agent Design
S5 Real APIs
S4 Tool Use
Tools → Real APIs → ReAct → RAG → Multi-Agent
Section 09 — Key Takeaways + Homework

What to Remember

Agent teams mirror org charts. Division of labor, specialization, and coordination costs apply to AI agents just like human teams. Design your agent system like you'd design a department.
Watch for the telephone game. Information degrades across agents. Citation chains and confidence propagation are your safeguards against compounding errors.
Multi-agent isn't always better. Each agent adds latency, cost, and complexity. Use the decision framework: multi-agent for complex, multi-expertise tasks. Single-agent for everything else.
The pipeline pattern is your workhorse. Research → Analysis → Review is the most common multi-agent pattern. Master it and you can adapt it to any domain.

Homework

Extend Add a 4th agent to your pipeline (e.g., a Fact-Checker, Summarizer, or Domain Expert). Test the full pipeline with 3 business questions.
Analyze Run your pipeline 5 times on the same question. Document how outputs differ, where distortion occurs, and what safeguards would help.
Design Propose a multi-agent system for your target industry (the one you'll use for your final project). Include: agent roles, handoff logic, human gates, and cost estimate.
Read Read about Claude Code's multi-agent "team" feature. Write a paragraph on how it applies the orchestrator pattern from Session 6.
Next: Session 9 — Agents for Your Industry
Next session is different. No more tutorials. No more following along. YOU pick the industry. YOU design the agent. YOU build it. Session 9 is where everything you've learned becomes YOUR competitive advantage.
Preview Session 9