Multi-Agent System Design Patterns: A Practical Guide

Pattern 1: Pipeline (Sequential Handoff)

The Pipeline pattern chains agents in a fixed sequence: Agent A completes its task, passes output to Agent B, which passes to Agent C. Each agent transforms the work product, like a factory assembly line. This is the most common and simplest multi-agent architecture pattern.

When to use: Content pipelines, document review workflows, staged data processing, ETL transformations, and any task where each step depends on the previous step completing successfully.

Concrete Example: A content marketing pipeline with four named agents. The Content Writer creates a draft from a brief. The Editor reviews for tone, clarity, and brand voice. The Fact-Checker verifies claims and adds citations. The Publisher formats for the target platform and schedules distribution.

{
  "pattern": "pipeline",
  "agents": [
    {
      "id": "writer",
      "name": "Content Writer",
      "role": "Create high-quality first drafts from content briefs",
      "model": "gpt-4o",
      "tools": ["web_search", "document_reader"],
      "passes_to": "editor",
      "output_format": "markdown_draft"
    },
    {
      "id": "editor",
      "name": "Editor",
      "role": "Review drafts for clarity, tone, and brand voice consistency",
      "model": "gpt-4o",
      "tools": ["document_reader"],
      "passes_to": "fact_checker",
      "output_format": "edited_draft"
    },
    {
      "id": "fact_checker",
      "name": "Fact Checker",
      "role": "Verify all claims, statistics, and named references",
      "model": "gpt-4o",
      "tools": ["web_search", "calculator"],
      "passes_to": "publisher",
      "output_format": "verified_draft"
    },
    {
      "id": "publisher",
      "name": "Publisher",
      "role": "Format content for target platforms and schedule distribution",
      "model": "gpt-4o-mini",
      "tools": ["content_management_system"],
      "output_format": "published_url"
    }
  ],
  "execution": {
    "type": "sequential",
    "error_handling": "retry_then_escalate",
    "max_retries": 2
  }
}

Pattern 2: Fan-out / Fan-in (Parallel + Aggregation)

The Fan-out/Fan-in pattern uses a coordinator agent to dispatch the same task (or variations of it) to multiple specialist agents running in parallel, then aggregates their responses. This is the go-to pattern for gathering diverse perspectives, competitive analysis, or research from multiple sources simultaneously.

When to use: Research gathering from multiple sources, competitive analysis, multi-perspective evaluation, sentiment analysis across reviews, or any task where getting answers from diverse angles improves quality.

Concrete Example: A competitive intelligence system. The Coordinator receives a product query and fans out to three agents: a Market Researcher (finds market size and trends), a Feature Analyst (compares feature matrices), and a User Sentiment Critic (summarizes review sentiment). Their outputs go to a Summarizer that produces the final report.

{
  "pattern": "fan_out_fan_in",
  "coordinator": {
    "id": "coordinator",
    "name": "Research Coordinator",
    "role": "Decompose research questions and dispatch to specialists",
    "model": "gpt-4o-mini",
    "dispatch_strategy": "same_task_parallel"
  },
  "parallel_agents": [
    {
      "id": "market_researcher",
      "name": "Market Researcher",
      "role": "Find market size data, growth trends, and funding activity",
      "model": "gpt-4o",
      "tools": ["web_search", "data_analysis"],
      "output_format": "market_brief"
    },
    {
      "id": "feature_analyst",
      "name": "Feature Analyst",
      "role": "Compare product features, pricing tiers, and positioning",
      "model": "gpt-4o",
      "tools": ["web_search", "document_reader"],
      "output_format": "comparison_table"
    },
    {
      "id": "sentiment_critic",
      "name": "Sentiment Critic",
      "role": "Analyze user reviews, community feedback, and social sentiment",
      "model": "gpt-4o",
      "tools": ["web_search"],
      "output_format": "sentiment_summary"
    }
  ],
  "aggregator": {
    "id": "summarizer",
    "name": "Research Summarizer",
    "role": "Synthesize agent outputs into a unified competitive brief",
    "model": "gpt-4o",
    "output_format": "final_report"
  },
  "execution": {
    "type": "parallel_dispatch",
    "timeout_per_agent": 60,
    "failure_mode": "partial_results_ok",
    "min_agents_required": 2
  }
}

Pattern 3: Orchestrator-Worker

The Orchestrator-Worker pattern places a central "brain" agent in charge of understanding the task, decomposing it into dynamic subtasks, and delegating each to a specialist worker. Unlike Fan-out (which gives all agents the same job), the Orchestrator creates different, tailored tasks for each worker based on the specific request.

When to use: Complex projects requiring dynamic task decomposition, software development tasks, multi-step problem solving where subtasks differ substantially, and scenarios where the orchestrator needs to adapt its plan based on early results.

Concrete Example: A software development system where the Project Manager agent receives a feature request, analyzes requirements, creates a task breakdown, and delegates: the API Developer builds endpoints, the Database Engineer designs the schema, and the Test Writer creates test cases. The Project Manager then assembles the final deliverable.

Pattern 4: Hierarchical (Manager-of-Managers)

The Hierarchical pattern nests multiple levels of orchestration — a top-level manager delegates to mid-level managers, who each manage their own teams of specialist agents. It's essentially Orchestrator-Worker applied recursively, mirroring how large organizations structure their work.

When to use: Large-scale projects requiring coordination across multiple functional areas, enterprise applications simulating organizational workflows, product launches with engineering + marketing + operations streams, and any scenario where you need 8+ agents organized into sub-teams.

Concrete Example: A product launch system. The VP agent receives the launch brief, delegates the engineering workstream to an Engineering Lead (who manages Frontend Dev, Backend Dev, and QA), the marketing workstream to a Marketing Lead (Content Writer, SEO Analyst, Designer), and operations to an Operations Lead (DevOps Engineer, Data Analyst). Each lead integrates their team's output and reports back to the VP.

Pattern 5: Debate / Deliberation

The Debate pattern (also called Deliberation or Adversarial Collaboration) has multiple agents argue about the best answer, challenge each other's reasoning, and converge on a refined conclusion through structured rounds of critique. A Judge agent evaluates arguments and declares a winner or synthesizes the best elements.

When to use: High-stakes decisions where hallucination is dangerous, complex reasoning tasks that benefit from adversarial checking, legal analysis, strategic planning, investment decisions, and any scenario where a single agent might confidently produce wrong answers.

Concrete Example: A strategic decision system for evaluating market entry. The Advocate agent builds the case for entering a new market. The Opponent agent identifies risks, counterarguments, and reasons not to enter. After two rounds of debate, the Judge agent evaluates both sides' strongest arguments, identifies the most defensible position, and produces a final recommendation with explicit risk acknowledgment.