Multi agent orchestration

Multi-Agent Orchestration

Patterns for coordinating multiple AI agents to solve complex tasks. Covers orchestrator, pipeline, consensus, delegation, supervisor, and swarm architectures.

When to Use

• User is building a system with multiple cooperating agents
• User needs task delegation or agent supervision patterns
• User wants consensus-based decision making across agents
• User is designing pipeline processing with agent stages
• User asks about swarm intelligence or emergent agent behavior

Core Patterns

Orchestrator Pattern

A central orchestrator decomposes tasks and delegates to specialized worker agents.

import anthropic

client = anthropic.Anthropic()

def orchestrator(task: str) -> str:
    # Step 1: Plan and decompose
    plan = client.messages.create(
        model="claude-sonnet-4-6-20250514",
        max_tokens=2048,
        system="""You are a task orchestrator. Break the task into subtasks.
Return a JSON array of subtasks, each with "id", "agent", "instruction", and "depends_on" (list of ids).
Available agents: researcher, coder, reviewer.""",
        messages=[{"role": "user", "content": task}]
    )

    subtasks = json.loads(plan.content[0].text)

    # Step 2: Execute subtasks respecting dependencies
    results = {}
    for subtask in topological_sort(subtasks):
        dep_context = "\n".join(
            f"Result of {d}: {results[d]}" for d in subtask["depends_on"]
        )
        result = run_worker(
            agent=subtask["agent"],
            instruction=subtask["instruction"],
            context=dep_context
        )
        results[subtask["id"]] = result

    # Step 3: Synthesize final result
    synthesis = client.messages.create(
        model="claude-sonnet-4-6-20250514",
        max_tokens=4096,
        system="Synthesize the worker results into a coherent final response.",
        messages=[{"role": "user", "content": json.dumps(results)}]
    )
    return synthesis.content[0].text

def run_worker(agent: str, instruction: str, context: str) -> str:
    system_prompts = {
        "researcher": "You are a research agent. Find and summarize relevant information.",
        "coder": "You are a coding agent. Write clean, tested code.",
        "reviewer": "You are a review agent. Find bugs, security issues, and improvements."
    }
    response = client.messages.create(
        model="claude-haiku-4-5-20251001",  # Workers use faster model
        max_tokens=2048,
        system=system_prompts[agent],
        messages=[{"role": "user", "content": f"{instruction}\n\nContext:\n{context}"}]
    )
    return response.content[0].text

Pipeline Pattern

Agents process data sequentially, each stage transforming the output for the next.

def pipeline(input_text: str) -> dict:
    stages = [
        ("extract", "Extract all entities, facts, and claims from this text. Return structured JSON."),
        ("validate", "Verify each fact and claim. Mark each as verified, unverified, or false. Return updated JSON."),
        ("summarize", "Create a concise summary highlighting only verified facts. Return final JSON with summary field.")
    ]

    current = input_text
    for stage_name, instruction in stages:
        response = client.messages.create(
            model="claude-sonnet-4-6-20250514",
            max_tokens=4096,
            system=f"You are the {stage_name} stage of a processing pipeline. {instruction}",
            messages=[{"role": "user", "content": current}]
        )
        current = response.content[0].text

    return json.loads(current)

Consensus Pattern

Multiple agents independently analyze the same input, then a judge resolves disagreements.

def consensus_review(code: str) -> dict:
    perspectives = [
        ("security_expert", "Review for security vulnerabilities. Rate severity."),
        ("performance_engineer", "Review for performance issues and optimization opportunities."),
        ("maintainability_reviewer", "Review for code quality, readability, and maintainability.")
    ]

    # Gather independent reviews in parallel
    reviews = {}
    for role, instruction in perspectives:
        response = client.messages.create(
            model="claude-sonnet-4-6-20250514",
            max_tokens=2048,
            system=f"You are a {role}. {instruction}",
            messages=[{"role": "user", "content": f"Review this code:\n```\n{code}\n```"}]
        )
        reviews[role] = response.content[0].text

    # Judge synthesizes and resolves conflicts
    judge_response = client.messages.create(
        model="claude-sonnet-4-6-20250514",
        max_tokens=4096,
        system="""You are a senior engineering judge. Synthesize multiple code reviews.
Resolve any disagreements. Produce a final verdict with prioritized action items.
Return JSON with: overall_rating, critical_issues, recommendations, and dissenting_opinions.""",
        messages=[{"role": "user", "content": json.dumps(reviews)}]
    )
    return json.loads(judge_response.content[0].text)

Delegation Pattern

An agent decides at runtime which specialist to delegate to.

def delegating_agent(user_request: str) -> str:
    # Agent decides which specialist to invoke
    routing = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=256,
        system="""Route the request to the best specialist. Return JSON:
{"specialist": "sql_expert|api_designer|frontend_dev|devops_engineer", "refined_task": "..."}""",
        messages=[{"role": "user", "content": user_request}]
    )

    route = json.loads(routing.content[0].text)

    specialist_prompts = {
        "sql_expert": "You write optimized, safe SQL queries. Always use parameterized queries.",
        "api_designer": "You design RESTful APIs following OpenAPI 3.0 best practices.",
        "frontend_dev": "You build accessible, performant React components.",
        "devops_engineer": "You write infrastructure as code and CI/CD pipelines."
    }

    result = client.messages.create(
        model="claude-sonnet-4-6-20250514",
        max_tokens=4096,
        system=specialist_prompts[route["specialist"]],
        messages=[{"role": "user", "content": route["refined_task"]}]
    )
    return result.content[0].text

Supervisor Pattern

A supervisor monitors worker agents, intervenes on failure, and ensures quality.

def supervised_execution(task: str, max_retries: int = 3) -> str:
    for attempt in range(max_retries):
        # Worker attempts the task
        worker_result = client.messages.create(
            model="claude-haiku-4-5-20251001",
            max_tokens=4096,
            system="Complete the task. Return your result in <result> tags and confidence (0-1) in <confidence> tags.",
            messages=[{"role": "user", "content": task}]
        )
        worker_output = worker_result.content[0].text

        # Supervisor evaluates quality
        evaluation = client.messages.create(
            model="claude-sonnet-4-6-20250514",
            max_tokens=1024,
            system="""Evaluate the worker's output. Return JSON:
{"approved": true/false, "issues": ["..."], "guidance": "feedback for retry if not approved"}""",
            messages=[{
                "role": "user",
                "content": f"Task: {task}\n\nWorker output:\n{worker_output}"
            }]
        )

        verdict = json.loads(evaluation.content[0].text)
        if verdict["approved"]:
            return worker_output

        # Provide feedback for next attempt
        task = f"{task}\n\nPrevious attempt feedback: {verdict['guidance']}"

    return worker_output  # Return best effort after max retries

Anti-Patterns

• Using the most expensive model for every agent (use Haiku for workers, Sonnet for orchestrators)
• Not passing context between dependent agents (each agent works blind)
• Running all agents sequentially when they could run in parallel
• Letting agents communicate in free-form text without structured interfaces
• No termination condition in agentic loops (infinite retries)
• Single agent doing everything instead of decomposing into specialists
• Not logging intermediate results (makes debugging impossible)

Quick Reference

Pattern	When to Use	Tradeoff
Orchestrator	Complex tasks needing decomposition	Flexible but adds latency
Pipeline	Sequential data transformation	Simple but rigid ordering
Consensus	High-stakes decisions needing validation	Thorough but expensive
Delegation	Variable task types needing routing	Fast but needs good routing
Supervisor	Quality-critical output needing review	Reliable but slower
Swarm	Emergent problem-solving	Adaptive but hard to debug

Model selection for agents:

• Orchestrator / Judge / Supervisor: claude-sonnet-4-6 or claude-opus-4-6
• Workers / Routers: claude-haiku-4-5 (3x cost savings)
• Critical analysis: claude-opus-4-6 with extended thinking