Swarms - Pytorch
Project description
The Enterprise-Grade Production-Ready Multi-Agent Orchestration Framework
🐦 Twitter • 📢 Discord • Swarms Platform • 📙 Documentation
Swarms is an enterprise grade and production ready multi-agent collaboration framework that enables you to orchestrate many agents to work collaboratively at scale to automate real-world activities.
Requirements
python3.10or above!$ pip install -U swarmsAnd, don't forget to install swarms!.envfile with API keys from your providers likeOPENAI_API_KEY,ANTHROPIC_API_KEY- Set an
.envVariable with your desired workspace dir:WORKSPACE_DIR="agent_workspace"or do it in your terminal withexport WORKSPACE_DIR="agent_workspace"
Onboarding
Refer to our documentation for production grade implementation details.
| Section | Links |
|---|---|
| Installation | Installation |
| Quickstart | Get Started |
| Agent Internal Mechanisms | Agent Architecture |
| Agent API | Agent API |
| Integrating External Agents Griptape, Autogen, etc | Integrating External APIs |
| Creating Agents from YAML | Creating Agents from YAML |
| Why You Need Swarms | Why MultiAgent Collaboration is Necessary |
| Swarm Architectures Analysis | Swarm Architectures |
| Choosing the Right Swarm for Your Business Problem¶ | CLICK HERE |
| AgentRearrange Docs | CLICK HERE |
Install 💻
$ pip3 install -U swarms
Onboarding
Now that you have downloaded swarms with pip3 install -U swarms, we get access to the CLI. Get Onboarded with CLI Now with:
swarms onboarding
You can also run this command for help:
swarms help
Output
_________
/ _____/_ _ _______ _______ _____ ______
\_____ \ \/ \/ /\__ \_ __ \/ \ / ___/
/ \ / / __ \| | \/ Y Y \___ \
/_______ / \/\_/ (____ /__| |__|_| /____ >
\/ \/ \/ \/
Swarms CLI - Help
Commands:
onboarding : Starts the onboarding process
help : Shows this help message
get-api-key : Retrieves your API key from the platform
check-login : Checks if you're logged in and starts the cache
read-docs : Redirects you to swarms cloud documentation!
run-agents : Run your Agents from your agents.yaml
For more details, visit: https://docs.swarms.world
For more documentation on the CLI CLICK HERE
Usage Examples 🤖
Here are some simple examples but we have more comprehensive documentation at our docs here
Agent Class
The Agent class is a fundamental component of the Swarms framework, designed to execute tasks autonomously. It fuses llms, tools and long-term memory capabilities to create a full stack agent. The Agent class is highly customizable, allowing for fine-grained control over its behavior and interactions.
run Method
The run method is the primary entry point for executing tasks with an Agent instance. It accepts a task string as the main input task and processes it according to the agent's configuration. And, it can also accept an img parameter such as img="image_filepath.png to process images if you have a VLM
Settings and Customization
The Agent class offers a range of settings to tailor its behavior to specific needs. Some key settings include:
| Setting | Description | Default Value |
|---|---|---|
agent_name |
The name of the agent. | "DefaultAgent" |
system_prompt |
The system prompt to use for the agent. | "Default system prompt." |
llm |
The language model to use for processing tasks. | OpenAIChat instance |
max_loops |
The maximum number of loops to execute for a task. | 1 |
autosave |
Enables or disables autosaving of the agent's state. | False |
dashboard |
Enables or disables the dashboard for the agent. | False |
verbose |
Controls the verbosity of the agent's output. | False |
dynamic_temperature_enabled |
Enables or disables dynamic temperature adjustment for the language model. | False |
saved_state_path |
The path to save the agent's state. | "agent_state.json" |
user_name |
The username associated with the agent. | "default_user" |
retry_attempts |
The number of retry attempts for failed tasks. | 1 |
context_length |
The maximum length of the context to consider for tasks. | 200000 |
return_step_meta |
Controls whether to return step metadata in the output. | False |
output_type |
The type of output to return (e.g., "json", "string"). | "string" |
import os
from swarms import Agent
from swarm_models import OpenAIChat
from swarms.prompts.finance_agent_sys_prompt import (
FINANCIAL_AGENT_SYS_PROMPT,
)
from dotenv import load_dotenv
load_dotenv()
# Get the OpenAI API key from the environment variable
api_key = os.getenv("OPENAI_API_KEY")
# Create an instance of the OpenAIChat class
model = OpenAIChat(
openai_api_key=api_key, model_name="gpt-4o-mini", temperature=0.1
)
# Initialize the agent
agent = Agent(
agent_name="Financial-Analysis-Agent",
system_prompt=FINANCIAL_AGENT_SYS_PROMPT,
llm=model,
max_loops=1,
autosave=True,
dashboard=False,
verbose=True,
dynamic_temperature_enabled=True,
saved_state_path="finance_agent.json",
user_name="swarms_corp",
retry_attempts=1,
context_length=200000,
return_step_meta=False,
# output_type="json",
)
out = agent.run(
"How can I establish a ROTH IRA to buy stocks and get a tax break? What are the criteria"
)
print(out)
Integrating RAG with Swarms for Enhanced Long-Term Memory
Agent equipped with quasi-infinite long term memory using RAG (Relational Agent Graph) for advanced document understanding, analysis, and retrieval capabilities.
Mermaid Diagram for RAG Integration
graph TD
A[Initialize Agent with RAG] --> B[Receive Task]
B --> C[Query Long-Term Memory]
C --> D[Process Task with Context]
D --> E[Generate Response]
E --> F[Update Long-Term Memory]
F --> G[Return Output]
Step 1: Initialize the ChromaDB Client
import os
from swarms_memory import ChromaDB
# Initialize the ChromaDB client for long-term memory management
chromadb = ChromaDB(
metric="cosine", # Metric for similarity measurement
output_dir="finance_agent_rag", # Directory for storing RAG data
# docs_folder="artifacts", # Uncomment and specify the folder containing your documents
)
Step 2: Define the Model
from swarm_models import Anthropic
from swarms.prompts.finance_agent_sys_prompt import (
FINANCIAL_AGENT_SYS_PROMPT,
)
# Define the Anthropic model for language processing
model = Anthropic(anthropic_api_key=os.getenv("ANTHROPIC_API_KEY"))
Step 3: Initialize the Agent with RAG
from swarms import Agent
# Initialize the agent with RAG capabilities
agent = Agent(
agent_name="Financial-Analysis-Agent",
system_prompt=FINANCIAL_AGENT_SYS_PROMPT,
agent_description="Agent creates a comprehensive financial analysis",
llm=model,
max_loops="auto", # Auto-adjusts loops based on task complexity
autosave=True, # Automatically saves agent state
dashboard=False, # Disables dashboard for this example
verbose=True, # Enables verbose mode for detailed output
streaming_on=True, # Enables streaming for real-time processing
dynamic_temperature_enabled=True, # Dynamically adjusts temperature for optimal performance
saved_state_path="finance_agent.json", # Path to save agent state
user_name="swarms_corp", # User name for the agent
retry_attempts=3, # Number of retry attempts for failed tasks
context_length=200000, # Maximum length of the context to consider
long_term_memory=chromadb, # Integrates ChromaDB for long-term memory management
)
# Run the agent with a sample task
agent.run(
"What are the components of a startups stock incentive equity plan"
)
Misc Agent Settings
We provide vast array of features to save agent states using json, yaml, toml, upload pdfs, batched jobs, and much more!
Method Table
| Method | Description |
|---|---|
to_dict() |
Converts the agent object to a dictionary. |
to_toml() |
Converts the agent object to a TOML string. |
model_dump_json() |
Dumps the model to a JSON file. |
model_dump_yaml() |
Dumps the model to a YAML file. |
ingest_docs() |
Ingests documents into the agent's knowledge base. |
receive_message() |
Receives a message from a user and processes it. |
send_agent_message() |
Sends a message from the agent to a user. |
filtered_run() |
Runs the agent with a filtered system prompt. |
bulk_run() |
Runs the agent with multiple system prompts. |
add_memory() |
Adds a memory to the agent. |
check_available_tokens() |
Checks the number of available tokens for the agent. |
tokens_checks() |
Performs token checks for the agent. |
print_dashboard() |
Prints the dashboard of the agent. |
get_docs_from_doc_folders() |
Fetches all the documents from the doc folders. |
activate_agentops() |
Activates agent operations. |
check_end_session_agentops() |
Checks the end of the session for agent operations. |
# # Convert the agent object to a dictionary
print(agent.to_dict())
print(agent.to_toml())
print(agent.model_dump_json())
print(agent.model_dump_yaml())
# Ingest documents into the agent's knowledge base
agent.ingest_docs("your_pdf_path.pdf")
# Receive a message from a user and process it
agent.receive_message(name="agent_name", message="message")
# Send a message from the agent to a user
agent.send_agent_message(agent_name="agent_name", message="message")
# Ingest multiple documents into the agent's knowledge base
agent.ingest_docs("your_pdf_path.pdf", "your_csv_path.csv")
# Run the agent with a filtered system prompt
agent.filtered_run(
"How can I establish a ROTH IRA to buy stocks and get a tax break? What are the criteria?"
)
# Run the agent with multiple system prompts
agent.bulk_run(
[
"How can I establish a ROTH IRA to buy stocks and get a tax break? What are the criteria?",
"Another system prompt",
]
)
# Add a memory to the agent
agent.add_memory("Add a memory to the agent")
# Check the number of available tokens for the agent
agent.check_available_tokens()
# Perform token checks for the agent
agent.tokens_checks()
# Print the dashboard of the agent
agent.print_dashboard()
# Fetch all the documents from the doc folders
agent.get_docs_from_doc_folders()
# Activate agent ops
agent.activate_agentops()
agent.check_end_session_agentops()
# Dump the model to a JSON file
agent.model_dump_json()
print(agent.to_toml())
Agentwith Pydantic BaseModel as Output Type
The following is an example of an agent that intakes a pydantic basemodel and outputs it at the same time:
from pydantic import BaseModel, Field
from swarms import Agent
from swarm_models import Anthropic
# Initialize the schema for the person's information
class Schema(BaseModel):
name: str = Field(..., title="Name of the person")
agent: int = Field(..., title="Age of the person")
is_student: bool = Field(..., title="Whether the person is a student")
courses: list[str] = Field(
..., title="List of courses the person is taking"
)
# Convert the schema to a JSON string
tool_schema = Schema(
name="Tool Name",
agent=1,
is_student=True,
courses=["Course1", "Course2"],
)
# Define the task to generate a person's information
task = "Generate a person's information based on the following schema:"
# Initialize the agent
agent = Agent(
agent_name="Person Information Generator",
system_prompt=(
"Generate a person's information based on the following schema:"
),
# Set the tool schema to the JSON string -- this is the key difference
tool_schema=tool_schema,
llm=Anthropic(),
max_loops=3,
autosave=True,
dashboard=False,
streaming_on=True,
verbose=True,
interactive=True,
# Set the output type to the tool schema which is a BaseModel
output_type=tool_schema, # or dict, or str
metadata_output_type="json",
# List of schemas that the agent can handle
list_base_models=[tool_schema],
function_calling_format_type="OpenAI",
function_calling_type="json", # or soon yaml
)
# Run the agent to generate the person's information
generated_data = agent.run(task)
# Print the generated data
print(f"Generated data: {generated_data}")
Multi Modal Autonomous Agent
Run the agent with multiple modalities useful for various real-world tasks in manufacturing, logistics, and health.
import os
from dotenv import load_dotenv
from swarms import Agent
from swarm_models import GPT4VisionAPI
# Load the environment variables
load_dotenv()
# Initialize the language model
llm = GPT4VisionAPI(
openai_api_key=os.environ.get("OPENAI_API_KEY"),
max_tokens=500,
)
# Initialize the task
task = (
"Analyze this image of an assembly line and identify any issues such as"
" misaligned parts, defects, or deviations from the standard assembly"
" process. IF there is anything unsafe in the image, explain why it is"
" unsafe and how it could be improved."
)
img = "assembly_line.jpg"
## Initialize the workflow
agent = Agent(
agent_name = "Multi-ModalAgent",
llm=llm,
max_loops="auto",
autosave=True,
dashboard=True,
multi_modal=True
)
# Run the workflow on a task
agent.run(task, img)
ToolAgent
ToolAgent is an agent that can use tools through JSON function calling. It intakes any open source model from huggingface and is extremely modular and plug in and play. We need help adding general support to all models soon.
from pydantic import BaseModel, Field
from transformers import AutoModelForCausalLM, AutoTokenizer
from swarms import ToolAgent
from swarms.utils.json_utils import base_model_to_json
# Load the pre-trained model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
"databricks/dolly-v2-12b",
load_in_4bit=True,
device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("databricks/dolly-v2-12b")
# Initialize the schema for the person's information
class Schema(BaseModel):
name: str = Field(..., title="Name of the person")
agent: int = Field(..., title="Age of the person")
is_student: bool = Field(
..., title="Whether the person is a student"
)
courses: list[str] = Field(
..., title="List of courses the person is taking"
)
# Convert the schema to a JSON string
tool_schema = base_model_to_json(Schema)
# Define the task to generate a person's information
task = (
"Generate a person's information based on the following schema:"
)
# Create an instance of the ToolAgent class
agent = ToolAgent(
name="dolly-function-agent",
description="Ana gent to create a child data",
model=model,
tokenizer=tokenizer,
json_schema=tool_schema,
)
# Run the agent to generate the person's information
generated_data = agent.run(task)
# Print the generated data
print(f"Generated data: {generated_data}")
Integrating External Agents
Integrating external agents from other agent frameworks is easy with swarms.
Steps:
- Create a new class that inherits
Agent - Create a
.run(task: str) -> strmethod that runs the agent and returns the response. - The new Agent must return a string of the response. But you may add additional methods to save the output to JSON.
Griptape Example
For example, here's an example on how to create an agent from griptape.
Here’s how you can create a custom Griptape agent that integrates with the Swarms framework by inheriting from the Agent class in Swarms and overriding the run(task: str) -> str method.
from swarms import (
Agent as SwarmsAgent,
) # Import the base Agent class from Swarms
from griptape.structures import Agent as GriptapeAgent
from griptape.tools import (
WebScraperTool,
FileManagerTool,
PromptSummaryTool,
)
# Create a custom agent class that inherits from SwarmsAgent
class GriptapeSwarmsAgent(SwarmsAgent):
def __init__(self, *args, **kwargs):
# Initialize the Griptape agent with its tools
self.agent = GriptapeAgent(
input="Load {{ args[0] }}, summarize it, and store it in a file called {{ args[1] }}.",
tools=[
WebScraperTool(off_prompt=True),
PromptSummaryTool(off_prompt=True),
FileManagerTool(),
],
*args,
**kwargs,
# Add additional settings
)
# Override the run method to take a task and execute it using the Griptape agent
def run(self, task: str) -> str:
# Extract URL and filename from task (you can modify this parsing based on task structure)
url, filename = task.split(
","
) # Example of splitting task string
# Execute the Griptape agent with the task inputs
result = self.agent.run(url.strip(), filename.strip())
# Return the final result as a string
return str(result)
# Example usage:
griptape_swarms_agent = GriptapeSwarmsAgent()
output = griptape_swarms_agent.run(
"https://griptape.ai, griptape.txt"
)
print(output)
Key Components:
- GriptapeSwarmsAgent: A custom class that inherits from the
SwarmsAgentclass and integrates the Griptape agent. - run(task: str) -> str: A method that takes a task string, processes it (e.g., splitting into a URL and filename), and runs the Griptape agent with the provided inputs.
- Griptape Tools: The tools integrated into the Griptape agent (e.g.,
WebScraperTool,PromptSummaryTool,FileManagerTool) allow for web scraping, summarization, and file management.
You can now easily plug this custom Griptape agent into the Swarms Framework and use it to run tasks!
Understanding Swarms
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:
-
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.
-
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.
-
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.
-
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.
-
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 |
SequentialWorkflow
Sequential Workflow enables you to sequentially execute tasks with Agent and then pass the output into the next agent and onwards until you have specified your max loops.
graph LR
A[Agent 1] --> B[Agent 2]
B --> C[Agent 3]
C --> D[Agent 4]
D --> E[Max Loops]
E --> F[End]
In this example, each Agent represents a task that is executed sequentially. The output of each agent is passed to the next agent in the sequence until the maximum number of loops is reached. This workflow is particularly useful for tasks that require a series of steps to be executed in a specific order, such as data processing pipelines or complex calculations that rely on the output of previous steps.
from swarms import Agent, SequentialWorkflow
from swarm_models import Anthropic
# Initialize the language model agent (e.g., GPT-3)
llm = Anthropic()
# Initialize agents for individual tasks
agent1 = Agent(
agent_name="Blog generator",
system_prompt="Generate a blog post like stephen king",
llm=llm,
max_loops=1,
dashboard=False,
tools=[],
)
agent2 = Agent(
agent_name="summarizer",
system_prompt="Sumamrize the blog post",
llm=llm,
max_loops=1,
dashboard=False,
tools=[],
)
# Create the Sequential workflow
workflow = SequentialWorkflow(
agents=[agent1, agent2], max_loops=1, verbose=False
)
# Run the workflow
workflow.run(
"Generate a blog post on how swarms of agents can help businesses grow."
)
AgentRearrange
The AgentRearrange orchestration technique, inspired by Einops and einsum, allows you to define and map out the relationships between various agents. It provides a powerful tool for orchestrating complex workflows, enabling you to specify linear and sequential relationships such as a -> a1 -> a2 -> a3, or concurrent relationships where the first agent sends a message to 3 agents simultaneously: a -> a1, a2, a3. This level of customization allows for the creation of highly efficient and dynamic workflows, where agents can work in parallel or in sequence as needed. The AgentRearrange technique is a valuable addition to the swarms library, providing a new level of flexibility and control over the orchestration of agents. For more detailed information and examples, please refer to the official documentation.
from swarms import Agent, AgentRearrange
from swarm_models import Anthropic
# Initialize the director agent
director = Agent(
agent_name="Director",
system_prompt="Directs the tasks for the workers",
llm=Anthropic(),
max_loops=1,
dashboard=False,
streaming_on=True,
verbose=True,
stopping_token="<DONE>",
state_save_file_type="json",
saved_state_path="director.json",
)
# Initialize worker 1
worker1 = Agent(
agent_name="Worker1",
system_prompt="Generates a transcript for a youtube video on what swarms are",
llm=Anthropic(),
max_loops=1,
dashboard=False,
streaming_on=True,
verbose=True,
stopping_token="<DONE>",
state_save_file_type="json",
saved_state_path="worker1.json",
)
# Initialize worker 2
worker2 = Agent(
agent_name="Worker2",
system_prompt="Summarizes the transcript generated by Worker1",
llm=Anthropic(),
max_loops=1,
dashboard=False,
streaming_on=True,
verbose=True,
stopping_token="<DONE>",
state_save_file_type="json",
saved_state_path="worker2.json",
)
# Create a list of agents
agents = [director, worker1, worker2]
# Define the flow pattern
flow = "Director -> Worker1 -> Worker2"
# Using AgentRearrange class
agent_system = AgentRearrange(agents=agents, flow=flow)
output = agent_system.run(
"Create a format to express and communicate swarms of llms in a structured manner for youtube"
)
print(output)
HierarhicalSwarm
Coming soon...
GraphSwarm
The GraphSwarm is a workflow management system designed to orchestrate complex tasks by leveraging the power of graph theory. It enables the creation of a directed acyclic graph (DAG) to model dependencies between tasks and agents. This allows for efficient task assignment, execution, and monitoring.
Here's a breakdown of how the GraphSwarm works:
- Node Creation: The
GraphSwarmworkflow is composed of nodes, which can be either agents or tasks. Agents are responsible for executing tasks, and tasks represent specific operations that need to be performed. In the example, two agents (agent1andagent2) and one task (task1) are created. - Edge Definition: Edges are used to define the relationships between nodes. In this case, edges are created to connect
agent1andagent2totask1, indicating that both agents are capable of executingtask1. - Entry and End Points: The
GraphSwarmworkflow requires the definition of entry points (where the workflow starts) and end points (where the workflow concludes). In this example,agent1andagent2are set as entry points, andtask1is set as the end point. - Visualization: The
GraphSwarmprovides a visualization feature to graphically represent the workflow. This allows for easy understanding and debugging of the workflow structure. - Execution: The
GraphSwarmworkflow is executed by traversing the graph from the entry points to the end points. In this case, bothagent1andagent2executetask1concurrently, and the results are collected. - Results: The final results of the workflow execution are aggregated and returned. In this example, the result of executing
task1is "Task completed".
The GraphSwarm offers several benefits, including:
- Concurrency: Enables the execution of tasks concurrently, improving overall workflow efficiency.
- Flexibility: Allows for dynamic task assignment based on agent availability and task requirements.
- Scalability: Supports the addition of new agents and tasks as needed, making it suitable for large-scale workflows.
- Visualization: Provides a graphical representation of the workflow, facilitating understanding and debugging.
By leveraging the GraphSwarm, complex workflows can be efficiently managed, and tasks can be executed in a coordinated and scalable manner.
import os
from dotenv import load_dotenv
from swarms import Agent, Edge, GraphWorkflow, Node, NodeType
from swarm_models import OpenAIChat
load_dotenv()
api_key = os.environ.get("OPENAI_API_KEY")
llm = OpenAIChat(
temperature=0.5, openai_api_key=api_key, max_tokens=4000
)
agent1 = Agent(llm=llm, max_loops=1, autosave=True, dashboard=True)
agent2 = Agent(llm=llm, max_loops=1, autosave=True, dashboard=True)
def sample_task():
print("Running sample task")
return "Task completed"
wf_graph = GraphWorkflow()
wf_graph.add_node(Node(id="agent1", type=NodeType.AGENT, agent=agent1))
wf_graph.add_node(Node(id="agent2", type=NodeType.AGENT, agent=agent2))
wf_graph.add_node(
Node(id="task1", type=NodeType.TASK, callable=sample_task)
)
wf_graph.add_edge(Edge(source="agent1", target="task1"))
wf_graph.add_edge(Edge(source="agent2", target="task1"))
wf_graph.set_entry_points(["agent1", "agent2"])
wf_graph.set_end_points(["task1"])
print(wf_graph.visualize())
# Run the workflow
results = wf_graph.run()
print("Execution results:", results)
MixtureOfAgents
This is an implementation based on the paper: "Mixture-of-Agents Enhances Large Language Model Capabilities" by together.ai, available at https://arxiv.org/abs/2406.04692. It achieves state-of-the-art (SOTA) results on AlpacaEval 2.0, MT-Bench, and FLASK, surpassing GPT-4 Omni. This architecture is particularly suitable for tasks that require parallelization followed by sequential processing in another loop.
from swarms import Agent, OpenAIChat, MixtureOfAgents
# Initialize the director agent
director = Agent(
agent_name="Director",
system_prompt="Directs the tasks for the accountants",
llm=OpenAIChat(),
max_loops=1,
dashboard=False,
streaming_on=True,
verbose=True,
stopping_token="<DONE>",
state_save_file_type="json",
saved_state_path="director.json",
)
# Initialize accountant 1
accountant1 = Agent(
agent_name="Accountant1",
system_prompt="Prepares financial statements",
llm=OpenAIChat(),
max_loops=1,
dashboard=False,
streaming_on=True,
verbose=True,
stopping_token="<DONE>",
state_save_file_type="json",
saved_state_path="accountant1.json",
)
# Initialize accountant 2
accountant2 = Agent(
agent_name="Accountant2",
system_prompt="Audits financial records",
llm=OpenAIChat(),
max_loops=1,
dashboard=False,
streaming_on=True,
verbose=True,
stopping_token="<DONE>",
state_save_file_type="json",
saved_state_path="accountant2.json",
)
# Create a list of agents
agents = [director, accountant1, accountant2]
# Swarm
swarm = MixtureOfAgents(
name="Mixture of Accountants",
agents=agents,
layers=3,
final_agent=director,
)
# Run the swarm
out = swarm.run("Prepare financial statements and audit financial records")
print(out)
SpreadSheetSwarm
A revolutionary swarm architecture designed for concurrent management and oversight of thousands of agents, facilitating a one-to-many approach for efficient task processing and output analysis.
Key Features
- Concurrency: Enables the simultaneous execution of multiple agents, significantly reducing processing time and increasing overall system efficiency.
- One-to-Many: Allows a single task to be dynamically distributed among multiple agents, ensuring that each agent is utilized to its full potential.
- Scalability: Supports the integration of thousands of agents, making it an ideal solution for large-scale task processing and data analysis.
import os
from swarms import Agent
from swarm_models import OpenAIChat
from swarms.structs.spreadsheet_swarm import SpreadSheetSwarm
# Define custom system prompts for each social media platform
TWITTER_AGENT_SYS_PROMPT = """
You are a Twitter marketing expert specializing in real estate. Your task is to create engaging, concise tweets to promote properties, analyze trends to maximize engagement, and use appropriate hashtags and timing to reach potential buyers.
"""
INSTAGRAM_AGENT_SYS_PROMPT = """
You are an Instagram marketing expert focusing on real estate. Your task is to create visually appealing posts with engaging captions and hashtags to showcase properties, targeting specific demographics interested in real estate.
"""
FACEBOOK_AGENT_SYS_PROMPT = """
You are a Facebook marketing expert for real estate. Your task is to craft posts optimized for engagement and reach on Facebook, including using images, links, and targeted messaging to attract potential property buyers.
"""
LINKEDIN_AGENT_SYS_PROMPT = """
You are a LinkedIn marketing expert for the real estate industry. Your task is to create professional and informative posts, highlighting property features, market trends, and investment opportunities, tailored to professionals and investors.
"""
EMAIL_AGENT_SYS_PROMPT = """
You are an Email marketing expert specializing in real estate. Your task is to write compelling email campaigns to promote properties, focusing on personalization, subject lines, and effective call-to-action strategies to drive conversions.
"""
# Example usage:
api_key = os.getenv("OPENAI_API_KEY")
# Model
model = OpenAIChat(
openai_api_key=api_key, model_name="gpt-4o-mini", temperature=0.1
)
# Initialize your agents for different social media platforms
agents = [
Agent(
agent_name="Twitter-RealEstate-Agent",
system_prompt=TWITTER_AGENT_SYS_PROMPT,
llm=model,
max_loops=1,
dynamic_temperature_enabled=True,
saved_state_path="twitter_realestate_agent.json",
user_name="realestate_swarms",
retry_attempts=1,
),
Agent(
agent_name="Instagram-RealEstate-Agent",
system_prompt=INSTAGRAM_AGENT_SYS_PROMPT,
llm=model,
max_loops=1,
dynamic_temperature_enabled=True,
saved_state_path="instagram_realestate_agent.json",
user_name="realestate_swarms",
retry_attempts=1,
),
Agent(
agent_name="Facebook-RealEstate-Agent",
system_prompt=FACEBOOK_AGENT_SYS_PROMPT,
llm=model,
max_loops=1,
dynamic_temperature_enabled=True,
saved_state_path="facebook_realestate_agent.json",
user_name="realestate_swarms",
retry_attempts=1,
),
Agent(
agent_name="LinkedIn-RealEstate-Agent",
system_prompt=LINKEDIN_AGENT_SYS_PROMPT,
llm=model,
max_loops=1,
dynamic_temperature_enabled=True,
saved_state_path="linkedin_realestate_agent.json",
user_name="realestate_swarms",
retry_attempts=1,
),
Agent(
agent_name="Email-RealEstate-Agent",
system_prompt=EMAIL_AGENT_SYS_PROMPT,
llm=model,
max_loops=1,
dynamic_temperature_enabled=True,
saved_state_path="email_realestate_agent.json",
user_name="realestate_swarms",
retry_attempts=1,
),
]
# Create a Swarm with the list of agents
swarm = SpreadSheetSwarm(
name="Real-Estate-Marketing-Swarm",
description="A swarm that processes real estate marketing tasks using multiple agents on different threads.",
agents=agents,
autosave_on=True,
save_file_path="real_estate_marketing_spreadsheet.csv",
run_all_agents=False,
max_loops=2,
)
# Run the swarm
swarm.run(
task="""
Create posts to promote luxury properties in North Texas, highlighting their features, location, and investment potential. Include relevant hashtags, images, and engaging captions.
Property:
$10,399,000
1609 Meandering Way Dr, Roanoke, TX 76262
Link to the property: https://www.zillow.com/homedetails/1609-Meandering-Way-Dr-Roanoke-TX-76262/308879785_zpid/
What's special
Unveiling a new custom estate in the prestigious gated Quail Hollow Estates! This impeccable residence, set on a sprawling acre surrounded by majestic trees, features a gourmet kitchen equipped with top-tier Subzero and Wolf appliances. European soft-close cabinets and drawers, paired with a double Cambria Quartzite island, perfect for family gatherings. The first-floor game room&media room add extra layers of entertainment. Step into the outdoor sanctuary, where a sparkling pool and spa, and sunken fire pit, beckon leisure. The lavish master suite features stunning marble accents, custom his&her closets, and a secure storm shelter.Throughout the home,indulge in the visual charm of designer lighting and wallpaper, elevating every space. The property is complete with a 6-car garage and a sports court, catering to the preferences of basketball or pickleball enthusiasts. This residence seamlessly combines luxury&recreational amenities, making it a must-see for the discerning buyer.
Facts & features
Interior
Bedrooms & bathrooms
Bedrooms: 6
Bathrooms: 8
Full bathrooms: 7
1/2 bathrooms: 1
Primary bedroom
Bedroom
Features: Built-in Features, En Suite Bathroom, Walk-In Closet(s)
Cooling
Central Air, Ceiling Fan(s), Electric
Appliances
Included: Built-In Gas Range, Built-In Refrigerator, Double Oven, Dishwasher, Gas Cooktop, Disposal, Ice Maker, Microwave, Range, Refrigerator, Some Commercial Grade, Vented Exhaust Fan, Warming Drawer, Wine Cooler
Features
Wet Bar, Built-in Features, Dry Bar, Decorative/Designer Lighting Fixtures, Eat-in Kitchen, Elevator, High Speed Internet, Kitchen Island, Pantry, Smart Home, Cable TV, Walk-In Closet(s), Wired for Sound
Flooring: Hardwood
Has basement: No
Number of fireplaces: 3
Fireplace features: Living Room, Primary Bedroom
Interior area
Total interior livable area: 10,466 sqft
Total spaces: 12
Parking features: Additional Parking
Attached garage spaces: 6
Carport spaces: 6
Features
Levels: Two
Stories: 2
Patio & porch: Covered
Exterior features: Built-in Barbecue, Barbecue, Gas Grill, Lighting, Outdoor Grill, Outdoor Living Area, Private Yard, Sport Court, Fire Pit
Pool features: Heated, In Ground, Pool, Pool/Spa Combo
Fencing: Wrought Iron
Lot
Size: 1.05 Acres
Details
Additional structures: Outdoor Kitchen
Parcel number: 42232692
Special conditions: Standard
Construction
Type & style
Home type: SingleFamily
Architectural style: Contemporary/Modern,Detached
Property subtype: Single Family Residence
"""
)
ForestSwarm
The ForestSwarm architecture is designed for efficient task assignment by dynamically selecting the most suitable agent from a collection of trees. This is achieved through asynchronous task processing, where agents are chosen based on their relevance to the task at hand. The relevance is determined by calculating the similarity between the system prompts associated with each agent and the keywords present in the task itself. For a more in-depth understanding of how ForestSwarm works, please refer to the official documentation.
from swarms.structs.tree_swarm import TreeAgent, Tree, ForestSwarm
# Example Usage:
# Create agents with varying system prompts and dynamically generated distances/keywords
agents_tree1 = [
TreeAgent(
system_prompt="Stock Analysis Agent",
agent_name="Stock Analysis Agent",
),
TreeAgent(
system_prompt="Financial Planning Agent",
agent_name="Financial Planning Agent",
),
TreeAgent(
agent_name="Retirement Strategy Agent",
system_prompt="Retirement Strategy Agent",
),
]
agents_tree2 = [
TreeAgent(
system_prompt="Tax Filing Agent",
agent_name="Tax Filing Agent",
),
TreeAgent(
system_prompt="Investment Strategy Agent",
agent_name="Investment Strategy Agent",
),
TreeAgent(
system_prompt="ROTH IRA Agent", agent_name="ROTH IRA Agent"
),
]
# Create trees
tree1 = Tree(tree_name="Financial Tree", agents=agents_tree1)
tree2 = Tree(tree_name="Investment Tree", agents=agents_tree2)
# Create the ForestSwarm
multi_agent_structure = ForestSwarm(trees=[tree1, tree2])
# Run a task
task = "Our company is incorporated in delaware, how do we do our taxes for free?"
output = multi_agent_structure.run(task)
print(output)
Onboarding Session
Get onboarded now with the creator and lead maintainer of Swarms, Kye Gomez, who will show you how to get started with the installation, usage examples, and starting to build your custom use case! CLICK HERE
Documentation
Documentation is located here at: docs.swarms.world
Folder Structure
The swarms package has been meticlously crafted for extreme use-ability and understanding, the swarms package is split up into various modules such as swarms.agents that holds pre-built agents, swarms.structs that holds a vast array of structures like Agent and multi agent structures. The 3 most important are structs, models, and agents.
├── __init__.py
├── agents
├── artifacts
├── memory
├── schemas
├── models -> swarm_models
├── prompts
├── structs
├── telemetry
├── tools
├── utils
└── workers
🫶 Contributions:
The easiest way to contribute is to pick any issue with the good first issue tag 💪. Read the Contributing guidelines here. Bug Report? File here | Feature Request? File here
Swarms is an open-source project, and contributions are VERY welcome. If you want to contribute, you can create new features, fix bugs, or improve the infrastructure. Please refer to the CONTRIBUTING.md and our contributing board to participate in Roadmap discussions!
Accelerate Backlog
Accelerate Bugs, Features, and Demos to implement by supporting us here:
Community
Join our growing community around the world, for real-time support, ideas, and discussions on Swarms 😊
- View our official Blog
- Chat live with us on Discord
- Follow us on Twitter
- Connect with us on LinkedIn
- Visit us on YouTube
- Join the Swarms community on Discord!
- Join our Swarms Community Gathering every Thursday at 1pm NYC Time to unlock the potential of autonomous agents in automating your daily tasks Sign up here
License
Creative Commons Attribution 4.0 International Public License
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
