multi-agent-patterns

name	multi-agent-patterns
description	Design multi-agent architectures for complex tasks. Use when single-agent context limits are exceeded, when tasks decompose naturally into subtasks, or when specializing agents improves quality. Includes OpenAI Agents SDK patterns (handoffs, as_tool) required for OmniChat Phase 5. Covers Triage routing, context passing, and streaming specialist patterns.

Multi-Agent Architecture Patterns

Multi-agent architectures distribute work across multiple language model instances, each with its own context window. When designed well, this distribution enables capabilities beyond single-agent limits. When designed poorly, it introduces coordination overhead that negates benefits. The critical insight is that sub-agents exist primarily to isolate context, not to anthropomorphize role division.

When to Activate

Activate this skill when:

• Single-agent context limits constrain task complexity
• Tasks decompose naturally into parallel subtasks
• Different subtasks require different tool sets or system prompts
• Building systems that must handle multiple domains simultaneously
• Scaling agent capabilities beyond single-context limits
• Designing production agent systems with multiple specialized components

Core Concepts

Multi-agent systems address single-agent context limitations through distribution. Three dominant patterns exist: supervisor/orchestrator for centralized control, peer-to-peer/swarm for flexible handoffs, and hierarchical for layered abstraction. The critical design principle is context isolation—sub-agents exist primarily to partition context rather than to simulate organizational roles.

Effective multi-agent systems require explicit coordination protocols, consensus mechanisms that avoid sycophancy, and careful attention to failure modes including bottlenecks, divergence, and error propagation.

Detailed Topics

Why Multi-Agent Architectures

The Context Bottleneck Single agents face inherent ceilings in reasoning capability, context management, and tool coordination. As tasks grow more complex, context windows fill with accumulated history, retrieved documents, and tool outputs. Performance degrades according to predictable patterns: the lost-in-middle effect, attention scarcity, and context poisoning.

Multi-agent architectures address these limitations by partitioning work across multiple context windows. Each agent operates in a clean context focused on its subtask. Results aggregate at a coordination layer without any single context bearing the full burden.

The Token Economics Reality Multi-agent systems consume significantly more tokens than single-agent approaches. Production data shows:

Architecture	Token Multiplier	Use Case
Single agent chat	1× baseline	Simple queries
Single agent with tools	~4× baseline	Tool-using tasks
Multi-agent system	~15× baseline	Complex research/coordination

Research on the BrowseComp evaluation found that three factors explain 95% of performance variance: token usage (80% of variance), number of tool calls, and model choice. This validates the multi-agent approach of distributing work across agents with separate context windows to add capacity for parallel reasoning.

Critically, upgrading to better models often provides larger performance gains than doubling token budgets. Claude Sonnet 4.5 showed larger gains than doubling tokens on earlier Sonnet versions. GPT-5.2's thinking mode similarly outperforms raw token increases. This suggests model selection and multi-agent architecture are complementary strategies.

The Parallelization Argument Many tasks contain parallelizable subtasks that a single agent must execute sequentially. A research task might require searching multiple independent sources, analyzing different documents, or comparing competing approaches. A single agent processes these sequentially, accumulating context with each step.

Multi-agent architectures assign each subtask to a dedicated agent with a fresh context. All agents work simultaneously, then return results to a coordinator. The total real-world time approaches the duration of the longest subtask rather than the sum of all subtasks.

The Specialization Argument Different tasks benefit from different agent configurations: different system prompts, different tool sets, different context structures. A general-purpose agent must carry all possible configurations in context. Specialized agents carry only what they need.

Multi-agent architectures enable specialization without combinatorial explosion. The coordinator routes to specialized agents; each agent operates with lean context optimized for its domain.

Architectural Patterns

Pattern 1: Supervisor/Orchestrator The supervisor pattern places a central agent in control, delegating to specialists and synthesizing results. The supervisor maintains global state and trajectory, decomposes user objectives into subtasks, and routes to appropriate workers.

User Query -> Supervisor -> [Specialist, Specialist, Specialist] -> Aggregation -> Final Output

When to use: Complex tasks with clear decomposition, tasks requiring coordination across domains, tasks where human oversight is important.

Advantages: Strict control over workflow, easier to implement human-in-the-loop interventions, ensures adherence to predefined plans.

Disadvantages: Supervisor context becomes bottleneck, supervisor failures cascade to all workers, "telephone game" problem where supervisors paraphrase sub-agent responses incorrectly.

The Telephone Game Problem and Solution LangGraph benchmarks found supervisor architectures initially performed 50% worse than optimized versions due to the "telephone game" problem where supervisors paraphrase sub-agent responses incorrectly, losing fidelity.

The fix: implement a forward_message tool allowing sub-agents to pass responses directly to users:

def forward_message(message: str, to_user: bool = True):
    """
    Forward sub-agent response directly to user without supervisor synthesis.
    
    Use when:
    - Sub-agent response is final and complete
    - Supervisor synthesis would lose important details
    - Response format must be preserved exactly
    """
    if to_user:
        return {"type": "direct_response", "content": message}
    return {"type": "supervisor_input", "content": message}

With this pattern, swarm architectures slightly outperform supervisors because sub-agents respond directly to users, eliminating translation errors.

Implementation note: Implement direct pass-through mechanisms allowing sub-agents to pass responses directly to users rather than through supervisor synthesis when appropriate.

Pattern 2: Peer-to-Peer/Swarm The peer-to-peer pattern removes central control, allowing agents to communicate directly based on predefined protocols. Any agent can transfer control to any other through explicit handoff mechanisms.

def transfer_to_agent_b():
    return agent_b  # Handoff via function return

agent_a = Agent(
    name="Agent A",
    functions=[transfer_to_agent_b]
)

When to use: Tasks requiring flexible exploration, tasks where rigid planning is counterproductive, tasks with emergent requirements that defy upfront decomposition.

Advantages: No single point of failure, scales effectively for breadth-first exploration, enables emergent problem-solving behaviors.

Disadvantages: Coordination complexity increases with agent count, risk of divergence without central state keeper, requires robust convergence constraints.

Implementation note: Define explicit handoff protocols with state passing. Ensure agents can communicate their context needs to receiving agents.

Pattern 3: Hierarchical Hierarchical structures organize agents into layers of abstraction: strategic, planning, and execution layers. Strategy layer agents define goals and constraints; planning layer agents break goals into actionable plans; execution layer agents perform atomic tasks.

Strategy Layer (Goal Definition) -> Planning Layer (Task Decomposition) -> Execution Layer (Atomic Tasks)

When to use: Large-scale projects with clear hierarchical structure, enterprise workflows with management layers, tasks requiring both high-level planning and detailed execution.

Advantages: Mirrors organizational structures, clear separation of concerns, enables different context structures at different levels.

Disadvantages: Coordination overhead between layers, potential for misalignment between strategy and execution, complex error propagation.

Context Isolation as Design Principle

The primary purpose of multi-agent architectures is context isolation. Each sub-agent operates in a clean context window focused on its subtask without carrying accumulated context from other subtasks.

Isolation Mechanisms Full context delegation: For complex tasks where the sub-agent needs complete understanding, the planner shares its entire context. The sub-agent has its own tools and instructions but receives full context for its decisions.

Instruction passing: For simple, well-defined subtasks, the planner creates instructions via function call. The sub-agent receives only the instructions needed for its specific task.

File system memory: For complex tasks requiring shared state, agents read and write to persistent storage. The file system serves as the coordination mechanism, avoiding context bloat from shared state passing.

Isolation Trade-offs Full context delegation provides maximum capability but defeats the purpose of sub-agents. Instruction passing maintains isolation but limits sub-agent flexibility. File system memory enables shared state without context passing but introduces latency and consistency challenges.

The right choice depends on task complexity, coordination needs, and acceptable latency.

Consensus and Coordination

The Voting Problem Simple majority voting treats hallucinations from weak models as equal to reasoning from strong models. Without intervention, multi-agent discussions devolve into consensus on false premises due to inherent bias toward agreement.

Weighted Voting Weight agent votes by confidence or expertise. Agents with higher confidence or domain expertise carry more weight in final decisions.

Debate Protocols Debate protocols require agents to critique each other's outputs over multiple rounds. Adversarial critique often yields higher accuracy on complex reasoning than collaborative consensus.

Trigger-Based Intervention Monitor multi-agent interactions for specific behavioral markers. Stall triggers activate when discussions make no progress. Sycophancy triggers detect when agents mimic each other's answers without unique reasoning.

Framework Considerations

Different frameworks implement these patterns with different philosophies. LangGraph uses graph-based state machines with explicit nodes and edges. AutoGen uses conversational/event-driven patterns with GroupChat. CrewAI uses role-based process flows with hierarchical crew structures.

OmniChat Constraint: Phase 5 (Chat Orchestrator) mandates the OpenAI Agents SDK for all agent logic. Do not introduce LangGraph, AutoGen, or CrewAI into the Orchestrator service.

OpenAI Agents SDK Patterns (OmniChat Required)

The OpenAI Agents SDK implements the Swarm pattern natively through the handoffs parameter. This is the mandatory approach for OmniChat.

Pattern: Triage Agent with Handoffs

The triage agent routes to specialists based on user intent. Each specialist has its own tools, instructions, and context isolation.

from agents import Agent, Runner

# Specialists with isolated contexts
search_agent = Agent(
    name="Search Specialist",
    handoff_description="Handles web search and information retrieval",
    instructions="Search the web and summarize findings.",
    tools=[web_search],
)

code_agent = Agent(
    name="Code Specialist",
    handoff_description="Handles code execution and debugging",
    instructions="Run Python code in sandbox and explain results.",
    tools=[run_python],
)

media_agent = Agent(
    name="Media Specialist",
    handoff_description="Handles image generation requests",
    instructions="Generate images from detailed prompts.",
    tools=[generate_image],
)

# Triage with handoffs (Swarm pattern)
triage_agent = Agent(
    name="Triage",
    instructions="""Analyze user requests and route to the appropriate specialist:
    - Search questions → Search Specialist
    - Code/programming → Code Specialist
    - Image requests → Media Specialist

    For general conversation, respond directly.""",
    handoffs=[search_agent, code_agent, media_agent],
)

# Runner automatically handles handoffs
result = await Runner.run(triage_agent, user_message)
print(result.last_agent.name)  # Shows which specialist handled it

Pattern: Context Passing Across Handoffs

Use RunContextWrapper to pass OmniChat-specific state (user_id, conversation_id) through agent transfers.

from dataclasses import dataclass
from agents import Agent, Runner, function_tool, RunContextWrapper

@dataclass
class OmniChatContext:
    user_id: str
    conversation_id: str
    message_id: str
    publisher: "EventPublisher"  # For streaming

@function_tool
async def web_search(
    wrapper: RunContextWrapper[OmniChatContext],
    query: str,
) -> str:
    """Search the web - publishes status to Stream Broker."""
    ctx = wrapper.context

    # Publish status update
    await ctx.publisher.status(ctx.user_id, "searching", f"Searching: {query}")

    # Execute search (via arq to Tool Executor)
    result = await execute_remote_search(query)

    return result

# Context flows through handoffs automatically
result = await Runner.run(
    triage_agent,
    user_message,
    context=OmniChatContext(
        user_id="user-123",
        conversation_id="conv-456",
        message_id="msg-789",
        publisher=event_publisher,
    ),
)

Pattern: Streaming Specialist (Solving the Telephone Game)

Sub-agents stream tokens directly to users via the Stream Broker, bypassing supervisor synthesis.

from agents import Agent, Runner
from agents.stream_events import RawResponsesStreamEvent

async def handle_streaming_chat(
    user_id: str,
    user_message: str,
    publisher: EventPublisher,
):
    """Stream specialist responses directly to user."""

    ctx = OmniChatContext(user_id=user_id, publisher=publisher, ...)

    # Publish initial status
    await publisher.status(user_id, "thinking", "Analyzing request...")

    # Run with streaming
    stream_result = Runner.run_streamed(triage_agent, user_message, context=ctx)

    async for event in stream_result.stream_events():
        if isinstance(event, RawResponsesStreamEvent):
            # Stream tokens directly to user (no supervisor paraphrasing)
            if hasattr(event.data, "delta") and event.data.delta.content:
                for part in event.data.delta.content:
                    if hasattr(part, "text"):
                        await publisher.delta(user_id, part.text, message_id)

    result = await stream_result.result()
    # result.last_agent tells us which specialist responded

This pattern solves the "Telephone Game Problem" because:

1. The specialist generates the response
2. Tokens stream directly to Stream Broker → SSE → User
3. Triage agent never rewrites or paraphrases the specialist's output

Pattern: Agents as Tools (Orchestration)

For complex workflows where the supervisor needs sub-agent outputs before continuing:

researcher = Agent(name="Researcher", instructions="Research topics thoroughly.")
analyst = Agent(name="Analyst", instructions="Analyze data and provide insights.")

orchestrator = Agent(
    name="Report Builder",
    instructions="""Build comprehensive reports:
    1. Use researcher to gather information
    2. Use analyst to analyze findings
    3. Synthesize into final report""",
    tools=[
        researcher.as_tool(
            tool_name="research",
            tool_description="Research a topic"
        ),
        analyst.as_tool(
            tool_name="analyze",
            tool_description="Analyze data"
        ),
    ],
)

Use this pattern when:

• Supervisor must see sub-agent results before proceeding
• Multiple sub-agent outputs must be combined
• Complex multi-step workflows with dependencies

Use handoffs pattern when:

• Specialist should respond directly to user
• Clean handoff without supervisor involvement
• Streaming responses are required

Practical Guidance

Failure Modes and Mitigations

Failure: Supervisor Bottleneck The supervisor accumulates context from all workers, becoming susceptible to saturation and degradation.

Mitigation: Implement output schema constraints so workers return only distilled summaries. Use checkpointing to persist supervisor state without carrying full history.

Failure: Coordination Overhead Agent communication consumes tokens and introduces latency. Complex coordination can negate parallelization benefits.

Mitigation: Minimize communication through clear handoff protocols. Batch results where possible. Use asynchronous communication patterns.

Failure: Divergence Agents pursuing different goals without central coordination can drift from intended objectives.

Mitigation: Define clear objective boundaries for each agent. Implement convergence checks that verify progress toward shared goals. Use time-to-live limits on agent execution.

Failure: Error Propagation Errors in one agent's output propagate to downstream agents that consume that output.

Mitigation: Validate agent outputs before passing to consumers. Implement retry logic with circuit breakers. Use idempotent operations where possible.

Examples

Example 1: Research Team Architecture

Supervisor
├── Researcher (web search, document retrieval)
├── Analyzer (data analysis, statistics)
├── Fact-checker (verification, validation)
└── Writer (report generation, formatting)

Example 2: Handoff Protocol

def handle_customer_request(request):
    if request.type == "billing":
        return transfer_to(billing_agent)
    elif request.type == "technical":
        return transfer_to(technical_agent)
    elif request.type == "sales":
        return transfer_to(sales_agent)
    else:
        return handle_general(request)

Guidelines

1. Design for context isolation as the primary benefit of multi-agent systems
2. Choose architecture pattern based on coordination needs, not organizational metaphor
3. Implement explicit handoff protocols with state passing
4. Use weighted voting or debate protocols for consensus
5. Monitor for supervisor bottlenecks and implement checkpointing
6. Validate outputs before passing between agents
7. Set time-to-live limits to prevent infinite loops
8. Test failure scenarios explicitly

Integration

This skill builds on context-fundamentals and context-degradation. It connects to:

• memory-systems - Shared state management across agents
• tool-design - Tool specialization per agent
• context-optimization - Context partitioning strategies

References

Internal reference:

• Frameworks Reference - Implementation patterns for all frameworks including OpenAI Agents SDK

OmniChat Skill Relationships

Skill	Role in Multi-Agent System
scaffolding-openai-agents	Syntactic guide for Agent/Runner/Tools
event-publisher	Publishes status/delta events during agent execution
streaming-llm-responses	Stream Broker consumes events and delivers via SSE
redis-pattern	arq task queue for remote tool dispatch

Framework Documentation

• OpenAI Agents SDK - Required for OmniChat Phase 5
• LangGraph Documentation - Alternative (not for OmniChat)
• AutoGen Framework - Alternative (not for OmniChat)
• CrewAI Documentation - Alternative (not for OmniChat)

---

Skill Metadata

Created: 2025-12-20 Last Updated: 2025-12-20 Author: Agent Skills for Context Engineering Contributors Version: 1.0.0