Skip to content

Swarm Architectures

What is a Swarm?

A swarm, in the context of multi-agent systems, refers to a group of more than two agents working collaboratively to achieve a common goal. These agents can be software entities, such as llms that interact with each other to perform complex tasks. The concept of a swarm is inspired by natural systems like ant colonies or bird flocks, where simple individual behaviors lead to complex group dynamics and problem-solving capabilities.

How Swarm Architectures Facilitate Communication

Swarm architectures are designed to establish and manage communication between agents within a swarm. These architectures define how agents interact, share information, and coordinate their actions to achieve the desired outcomes. Here are some key aspects of swarm architectures:

  1. Hierarchical Communication: In hierarchical swarms, communication flows from higher-level agents to lower-level agents. Higher-level agents act as coordinators, distributing tasks and aggregating results. This structure is efficient for tasks that require top-down control and decision-making.

  2. Parallel Communication: In parallel swarms, agents operate independently and communicate with each other as needed. This architecture is suitable for tasks that can be processed concurrently without dependencies, allowing for faster execution and scalability.

  3. Sequential Communication: Sequential swarms process tasks in a linear order, where each agent's output becomes the input for the next agent. This ensures that tasks with dependencies are handled in the correct sequence, maintaining the integrity of the workflow.

  4. Mesh Communication: In mesh swarms, agents are fully connected, allowing any agent to communicate with any other agent. This setup provides high flexibility and redundancy, making it ideal for complex systems requiring dynamic interactions.

  5. Federated Communication: Federated swarms involve multiple independent swarms that collaborate by sharing information and results. Each swarm operates autonomously but can contribute to a larger task, enabling distributed problem-solving across different nodes.

Swarm architectures leverage these communication patterns to ensure that agents work together efficiently, adapting to the specific requirements of the task at hand. By defining clear communication protocols and interaction models, swarm architectures enable the seamless orchestration of multiple agents, leading to enhanced performance and problem-solving capabilities.

Name Description Code Link Use Cases
Hierarchical Swarms A system where agents are organized in a hierarchy, with higher-level agents coordinating lower-level agents to achieve complex tasks. Code Link Manufacturing process optimization, multi-level sales management, healthcare resource coordination
Agent Rearrange A setup where agents rearrange themselves dynamically based on the task requirements and environmental conditions. Code Link Adaptive manufacturing lines, dynamic sales territory realignment, flexible healthcare staffing
Concurrent Workflows Agents perform different tasks simultaneously, coordinating to complete a larger goal. Code Link Concurrent production lines, parallel sales operations, simultaneous patient care processes
Sequential Coordination Agents perform tasks in a specific sequence, where the completion of one task triggers the start of the next. Code Link Step-by-step assembly lines, sequential sales processes, stepwise patient treatment workflows
Parallel Processing Agents work on different parts of a task simultaneously to speed up the overall process. Code Link Parallel data processing in manufacturing, simultaneous sales analytics, concurrent medical tests
Mixture of Agents A heterogeneous swarm where agents with different capabilities are combined to solve complex problems. Code Link Financial forecasting, complex problem-solving requiring diverse skills
Graph Workflow Agents collaborate in a directed acyclic graph (DAG) format to manage dependencies and parallel tasks. Code Link AI-driven software development pipelines, complex project management
Group Chat Agents engage in a chat-like interaction to reach decisions collaboratively. Code Link Real-time collaborative decision-making, contract negotiations
Agent Registry A centralized registry where agents are stored, retrieved, and invoked dynamically. Code Link Dynamic agent management, evolving recommendation engines
Spreadsheet Swarm Manages tasks at scale, tracking agent outputs in a structured format like CSV files. Code Link Large-scale marketing analytics, financial audits
Forest Swarm A swarm structure that organizes agents in a tree-like hierarchy for complex decision-making processes. Code Link Multi-stage workflows, hierarchical reinforcement learning
Swarm Router Routes and chooses the swarm architecture based on the task requirements and available agents. Code Link Dynamic task routing, adaptive swarm architecture selection, optimized agent allocation

Hierarchical Swarm

Overview: A Hierarchical Swarm architecture organizes the agents in a tree-like structure. Higher-level agents delegate tasks to lower-level agents, which can further divide tasks among themselves. This structure allows for efficient task distribution and scalability.

Use-Cases:

  • Complex decision-making processes where tasks can be broken down into subtasks.

  • Multi-stage workflows such as data processing pipelines or hierarchical reinforcement learning.

graph TD
    A[Root Agent] --> B1[Sub-Agent 1]
    A --> B2[Sub-Agent 2]
    B1 --> C1[Sub-Agent 1.1]
    B1 --> C2[Sub-Agent 1.2]
    B2 --> C3[Sub-Agent 2.1]
    B2 --> C4[Sub-Agent 2.2]

Parallel Swarm

Overview: In a Parallel Swarm architecture, multiple agents operate independently and simultaneously on different tasks. Each agent works on its own task without dependencies on the others. Learn more here in the docs:

Use-Cases:

  • Tasks that can be processed independently, such as parallel data analysis.

  • Large-scale simulations where multiple scenarios are run in parallel.

graph LR
    A[Task] --> B1[Sub-Agent 1]
    A --> B2[Sub-Agent 2]
    A --> B3[Sub-Agent 3]
    A --> B4[Sub-Agent 4]

Sequential Swarm

Overview: A Sequential Swarm architecture processes tasks in a linear sequence. Each agent completes its task before passing the result to the next agent in the chain. This architecture ensures orderly processing and is useful when tasks have dependencies. Learn more here in the docs:

Use-Cases:

  • Workflows where each step depends on the previous one, such as assembly lines or sequential data processing.

  • Scenarios requiring strict order of operations.

graph TD
    A[First Agent] --> B[Second Agent]
    B --> C[Third Agent]
    C --> D[Fourth Agent]

Round Robin Swarm

Overview: In a Round Robin Swarm architecture, tasks are distributed cyclically among a set of agents. Each agent takes turns handling tasks in a rotating order, ensuring even distribution of workload.

Use-Cases:

  • Load balancing in distributed systems.

  • Scenarios requiring fair distribution of tasks to avoid overloading any single agent.

graph TD
    A[Coordinator Agent] --> B1[Sub-Agent 1]
    A --> B2[Sub-Agent 2]
    A --> B3[Sub-Agent 3]
    A --> B4[Sub-Agent 4]
    B1 --> A
    B2 --> A
    B3 --> A
    B4 --> A

SpreadSheet Swarm

Overview: The SpreadSheet Swarm makes it easy to manage thousands of agents all in one place: a csv file. You can initialize any number of agents and then there is a loop parameter to run the loop of agents on the task. Learn more in the docs here

Use-Cases:

  • Multi-threaded execution: Execution agents on multiple threads

  • Save agent outputs into CSV file

  • One place to analyze agent outputs


graph TD
    A[Initialize SpreadSheetSwarm] --> B[Initialize Agents]
    B --> C[Load Task Queue]
    C --> D[Run Task]

    subgraph Agents
        D --> E1[Agent 1]
        D --> E2[Agent 2]
        D --> E3[Agent 3]
    end

    E1 --> F1[Process Task]
    E2 --> F2[Process Task]
    E3 --> F3[Process Task]

    F1 --> G1[Track Output]
    F2 --> G2[Track Output]
    F3 --> G3[Track Output]

    subgraph Save Outputs
        G1 --> H[Save to CSV]
        G2 --> H[Save to CSV]
        G3 --> H[Save to CSV]
    end

    H --> I{Autosave Enabled?}
    I --> |Yes| J[Export Metadata to JSON]
    I --> |No| K[End Swarm Run]

    %% Style adjustments
    classDef blackBox fill:#000,stroke:#f00,color:#fff;
    class A,B,C,D,E1,E2,E3,F1,F2,F3,G1,G2,G3,H,I,J,K blackBox;

Mixture of Agents Architecture


graph TD
    A[Task Input] --> B[Layer 1: Reference Agents]
    B --> C[Agent 1]
    B --> D[Agent 2]
    B --> E[Agent N]

    C --> F[Agent 1 Response]
    D --> G[Agent 2 Response]
    E --> H[Agent N Response]

    F & G & H --> I[Layer 2: Aggregator Agent]
    I --> J[Aggregate All Responses]
    J --> K[Final Output]

Alternative Experimental Architectures

1. Circular Swarm

Input Arguments:

  • name (str): Name of the swarm.
  • description (str): Description of the swarm.
  • goal (str): Goal of the swarm.
  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.
  • return_full_history (bool): Whether to return the full conversation history.

Functionality:

Agents pass tasks in a circular manner, where each agent works on the next task in the list.

graph TD
    Task1 --> Agent1
    Agent1 --> Agent2
    Agent2 --> Agent3
    Agent3 --> Task2
    Task2 --> Agent1

2. Linear Swarm

Input Arguments:

  • name (str): Name of the swarm.
  • description (str): Description of the swarm.
  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.
  • conversation (Conversation): Conversation object.
  • return_full_history (bool): Whether to return the full conversation history.

Functionality:

Agents pass tasks in a linear fashion, each agent working on one task sequentially.

graph LR
    Task1 --> Agent1
    Agent1 --> Agent2
    Agent2 --> Agent3
    Agent3 --> Task2

3. Star Swarm

Input Arguments:

  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.

Functionality:

A central agent (Agent 1) executes the tasks first, followed by the other agents working in parallel.

graph TD
    Task1 --> Agent1
    Agent1 --> Agent2
    Agent1 --> Agent3
    Agent1 --> Agent4

4. Mesh Swarm

Input Arguments:

  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.

Functionality:

Each agent works on tasks randomly from a task queue, until the task queue is empty.

graph TD
    Task1 --> Agent1
    Task2 --> Agent2
    Task3 --> Agent3
    Task4 --> Agent4
    Task5 --> Agent1
    Task6 --> Agent2

5. Grid Swarm

Input Arguments:

  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.

Functionality:

Agents are structured in a grid, and tasks are distributed accordingly.

graph TD
    Task1 --> Agent1
    Task2 --> Agent2
    Task3 --> Agent3
    Task4 --> Agent4

6. Pyramid Swarm

Input Arguments:

  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.

Functionality:

Agents are arranged in a pyramid structure. Each level of agents works in sequence.

graph TD
    Task1 --> Agent1
    Agent1 --> Agent2
    Agent2 --> Agent3
    Agent3 --> Task2

7. Fibonacci Swarm

Input Arguments:

  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.

Functionality:

Agents work according to the Fibonacci sequence, where the number of agents working on tasks follows this progression.

graph TD
    Task1 --> Agent1
    Agent1 --> Agent2
    Agent2 --> Agent3
    Task2 --> Agent5
    Agent5 --> Agent8

8. Prime Swarm

Input Arguments:

  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.

Functionality:

Agents are assigned tasks based on prime number indices in the list of agents.

graph TD
    Task1 --> Agent2
    Task2 --> Agent3
    Task3 --> Agent5
    Task4 --> Agent7

9. Power Swarm

Input Arguments:

  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.

Functionality:

Agents work on tasks following powers of two.

graph TD
    Task1 --> Agent1
    Task2 --> Agent2
    Task3 --> Agent4
    Task4 --> Agent8

10. Sigmoid Swarm

Input Arguments:

  • agents (AgentListType): List of agents involved.
  • tasks (List[str]): List of tasks for the agents.

Functionality:

Agents are selected based on the sigmoid function, with higher-indexed agents handling more complex tasks.

graph TD
    Task1 --> Agent1
    Task2 --> Agent2
    Task3 --> Agent3
    Task4 --> Agent4

11. Sinusoidal Swarm

Input Arguments:

  • agents (AgentListType): List of agents involved.
  • task (str): Task for the agents to work on.

Functionality:

Agents are assigned tasks based on a sinusoidal pattern.

graph TD
    Task --> Agent1
    Agent1 --> Agent2
    Agent2 --> Agent3
    Agent3 --> Task2

Each of these swarm architectures enables different task distribution and agent coordination strategies, making it possible to select the right architecture for specific types of agent-based problem-solving scenarios.

Examples

import asyncio
import os

from dotenv import load_dotenv
from loguru import logger
from swarm_models import OpenAIChat
from tickr_agent.main import TickrAgent

from swarms.structs.swarming_architectures import (
    circular_swarm,
    linear_swarm,
    mesh_swarm,
    pyramid_swarm,
    star_swarm,
)

# Load environment variables (API keys)
load_dotenv()
api_key = os.getenv("OPENAI_API_KEY")

# Initialize the OpenAI model
model = OpenAIChat(
    openai_api_key=api_key, model_name="gpt-4", temperature=0.1
)

# Custom Financial Agent System Prompts
STOCK_ANALYSIS_PROMPT = """
You are an expert financial analyst. Your task is to analyze stock market data for a company 
and provide insights on whether to buy, hold, or sell. Analyze trends, financial ratios, and market conditions.
"""

NEWS_SUMMARIZATION_PROMPT = """
You are a financial news expert. Summarize the latest news related to a company and provide insights on 
how it could impact its stock price. Be concise and focus on the key takeaways.
"""

RATIO_CALCULATION_PROMPT = """
You are a financial ratio analyst. Your task is to calculate key financial ratios for a company 
based on the available data, such as P/E ratio, debt-to-equity ratio, and return on equity. 
Explain what each ratio means for investors.
"""

# Example Usage
# Define stock tickers
stocks = ["AAPL", "TSLA"]


# Initialize Financial Analysis Agents
stock_analysis_agent = TickrAgent(
    agent_name="Stock-Analysis-Agent",
    system_prompt=STOCK_ANALYSIS_PROMPT,
    stocks=stocks,
)

news_summarization_agent = TickrAgent(
    agent_name="News-Summarization-Agent",
    system_prompt=NEWS_SUMMARIZATION_PROMPT,
    stocks=stocks,

)

ratio_calculation_agent = TickrAgent(
    agent_name="Ratio-Calculation-Agent",
    system_prompt=RATIO_CALCULATION_PROMPT,
    stocks=stocks,

)
# Create a list of agents for swarming
agents = [
    stock_analysis_agent,
    news_summarization_agent,
    ratio_calculation_agent,
]

# Define financial analysis tasks
tasks = [
    "Analyze the stock performance of Apple (AAPL) in the last 6 months.",
    "Summarize the latest financial news on Tesla (TSLA).",
    "Calculate the P/E ratio and debt-to-equity ratio for Amazon (AMZN).",
]

# -------------------------------# Showcase Circular Swarm
# -------------------------------
logger.info("Starting Circular Swarm for financial analysis.")
circular_result = circular_swarm(agents, tasks)
logger.info(f"Circular Swarm Result:\n{circular_result}\n")


# -------------------------------
# Showcase Linear Swarm
# -------------------------------
logger.info("Starting Linear Swarm for financial analysis.")
linear_result = linear_swarm(agents, tasks)
logger.info(f"Linear Swarm Result:\n{linear_result}\n")


# -------------------------------
# Showcase Star Swarm
# -------------------------------
logger.info("Starting Star Swarm for financial analysis.")
star_result = star_swarm(agents, tasks)
logger.info(f"Star Swarm Result:\n{star_result}\n")


# -------------------------------
# Showcase Mesh Swarm
# -------------------------------
logger.info("Starting Mesh Swarm for financial analysis.")
mesh_result = mesh_swarm(agents, tasks)
logger.info(f"Mesh Swarm Result:\n{mesh_result}\n")


# -------------------------------
# Showcase Pyramid Swarm
# -------------------------------
logger.info("Starting Pyramid Swarm for financial analysis.")
pyramid_result = pyramid_swarm(agents, tasks)
logger.info(f"Pyramid Swarm Result:\n{pyramid_result}\n")


# -------------------------------
# Example: One-to-One Communication between Agents
# -------------------------------
logger.info(
    "Starting One-to-One communication between Stock and News agents."
)
one_to_one_result = stock_analysis_agent.run(
    "Analyze Apple stock performance, and then send the result to the News Summarization Agent"
)
news_summary_result = news_summarization_agent.run(one_to_one_result)
logger.info(
    f"One-to-One Communication Result:\n{news_summary_result}\n"
)


# -------------------------------
# Example: Broadcasting to all agents
# -------------------------------
async def broadcast_task():
    logger.info("Broadcasting task to all agents.")
    task = "Summarize the overall stock market performance today."
    await asyncio.gather(*[agent.run(task) for agent in agents])


asyncio.run(broadcast_task())


# -------------------------------
# Deep Comments & Explanations
# -------------------------------

"""
Explanation of Key Components:

1. **Agents**:
   - We created three specialized agents for financial analysis: Stock Analysis, News Summarization, and Ratio Calculation.
   - Each agent is provided with a custom system prompt that defines their unique task in analyzing stock data.

2. **Swarm Examples**:
   - **Circular Swarm**: Agents take turns processing tasks in a circular manner.
   - **Linear Swarm**: Tasks are processed sequentially by each agent.
   - **Star Swarm**: The first agent (Stock Analysis) processes all tasks before distributing them to other agents.
   - **Mesh Swarm**: Agents work on random tasks from the task queue.
   - **Pyramid Swarm**: Agents are arranged in a pyramid structure, processing tasks layer by layer.

3. **One-to-One Communication**:
   - This showcases how one agent can pass its result to another agent for further processing, useful for complex workflows where agents depend on each other.

4. **Broadcasting**:
   - The broadcasting function demonstrates how a single task can be sent to all agents simultaneously. This can be useful for situations like summarizing daily stock market performance across multiple agents.

5. **Logging with Loguru**:
   - We use `loguru` for detailed logging throughout the swarms. This helps to track the flow of information and responses from each agent.
"""