Multi-Agent Orchestration is incredibly painful swarms is making it simple, seamless, and reliable.

🐦 Twitter   •   📢 Discord   •   Swarms Platform   •   📙 Documentation

", interactive=True, tools=[terminal, browser, file_editor, create_file], code_interpreter=True, # streaming=True, ) # Run the agent out = agent("Create a new file for a plan to take over the world.") print(out) ``` ### `Agent`with 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: ```python from pydantic import BaseModel, Field from swarms import Anthropic, Agent # 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_tool_schemas=[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. ```python # Description: This is an example of how to use the Agent class to run a multi-modal workflow import os from dotenv import load_dotenv from swarms import GPT4VisionAPI, Agent # Load the environment variables load_dotenv() # Get the API key from the environment api_key = os.environ.get("OPENAI_API_KEY") # Initialize the language model llm = GPT4VisionAPI( openai_api_key=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( llm=llm, max_loops="auto", autosave=True, dashboard=True, multi_modal=True ) # Run the workflow on a task agent.run(task=task, img=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. ```python 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}") ``` ### `Task` For deeper control of your agent stack, `Task` is a simple structure for task execution with the `Agent`. Imagine zapier like LLM-based workflow automation. ✅ Task is a structure for task execution with the Agent. ✅ Tasks can have descriptions, scheduling, triggers, actions, conditions, dependencies, priority, and a history. ✅ The Task structure allows for efficient workflow automation with LLM-based agents. ```python import os from dotenv import load_dotenv from swarms import Agent, OpenAIChat, Task # Load the environment variables load_dotenv() # Define a function to be used as the action def my_action(): print("Action executed") # Define a function to be used as the condition def my_condition(): print("Condition checked") return True # Create an agent agent = Agent( llm=OpenAIChat(openai_api_key=os.environ["OPENAI_API_KEY"]), max_loops=1, dashboard=False, ) # Create a task task = Task( description=( "Generate a report on the top 3 biggest expenses for small" " businesses and how businesses can save 20%" ), agent=agent, ) # Set the action and condition task.set_action(my_action) task.set_condition(my_condition) # Execute the task print("Executing task...") task.run() # Check if the task is completed if task.is_completed(): print("Task completed") else: print("Task not completed") # Output the result of the task print(f"Task result: {task.result}") ``` --- ---- # Multi-Agent Orchestration: Swarms was designed to facilitate the communication between many different and specialized agents from a vast array of other frameworks such as langchain, autogen, crew, and more. In traditional swarm theory, there are many types of swarms usually for very specialized use-cases and problem sets. Such as Hiearchical and sequential are great for accounting and sales, because there is usually a boss coordinator agent that distributes a workload to other specialized agents. | **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](https://docs.swarms.world/en/latest/swarms/structs/agent_rearrange/) | 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](https://docs.swarms.world/en/latest/swarms/structs/sequential_workflow/) | 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 | ### `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. `SequentialWorkflow` is wonderful for real-world business tasks like sending emails, summarizing documents, and analyzing data. ✅ Save and Restore Workflow states! ✅ Multi-Modal Support for Visual Chaining ✅ Utilizes Agent class ```python from swarms import Agent, SequentialWorkflow, 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." ) ``` ### `ConcurrentWorkflow` `ConcurrentWorkflow` runs all the tasks all at the same time with the inputs you give it! ```python import os from dotenv import load_dotenv from swarms import Agent, ConcurrentWorkflow, OpenAIChat, Task # Load environment variables from .env file load_dotenv() # Load environment variables llm = OpenAIChat(openai_api_key=os.getenv("OPENAI_API_KEY")) agent = Agent(llm=llm, max_loops=1) # Create a workflow workflow = ConcurrentWorkflow(max_workers=5) # Create tasks task1 = Task(agent, "What's the weather in miami") task2 = Task(agent, "What's the weather in new york") task3 = Task(agent, "What's the weather in london") # Add tasks to the workflow workflow.add(tasks=[task1, task2, task3]) # Run the workflow workflow.run() ``` ### `RecursiveWorkflow` `RecursiveWorkflow` will keep executing the tasks until a specific token like is located inside the text! ```python import os from dotenv import load_dotenv from swarms import Agent, OpenAIChat, RecursiveWorkflow, Task # Load environment variables from .env file load_dotenv() # Load environment variables llm = OpenAIChat(openai_api_key=os.getenv("OPENAI_API_KEY")) agent = Agent(llm=llm, max_loops=1) # Create a workflow workflow = RecursiveWorkflow(stop_token="") # Create tasks task1 = Task(agent, "What's the weather in miami") task2 = Task(agent, "What's the weather in new york") task3 = Task(agent, "What's the weather in london") # Add tasks to the workflow workflow.add(task1) workflow.add(task2) workflow.add(task3) # Run the workflow workflow.run() ``` ### `SwarmNetwork` `SwarmNetwork` provides the infrasturcture for building extremely dense and complex multi-agent applications that span across various types of agents. ✅ Efficient Task Management: SwarmNetwork's intelligent agent pool and task queue management system ensures tasks are distributed evenly across agents. This leads to efficient use of resources and faster task completion. ✅ Scalability: SwarmNetwork can dynamically scale the number of agents based on the number of pending tasks. This means it can handle an increase in workload by adding more agents, and conserve resources when the workload is low by reducing the number of agents. ✅ Versatile Deployment Options: With SwarmNetwork, each agent can be run on its own thread, process, container, machine, or even cluster. This provides a high degree of flexibility and allows for deployment that best suits the user's needs and infrastructure. ```python import os from dotenv import load_dotenv # Import the OpenAIChat model and the Agent struct from swarms import Agent, OpenAIChat, SwarmNetwork # Load the environment variables load_dotenv() # Get the API key from the environment api_key = os.environ.get("OPENAI_API_KEY") # Initialize the language model llm = OpenAIChat( temperature=0.5, openai_api_key=api_key, ) ## Initialize the workflow agent = Agent(llm=llm, max_loops=1, agent_name="Social Media Manager") agent2 = Agent(llm=llm, max_loops=1, agent_name=" Product Manager") agent3 = Agent(llm=llm, max_loops=1, agent_name="SEO Manager") # Load the swarmnet with the agents swarmnet = SwarmNetwork( agents=[agent, agent2, agent3], ) # List the agents in the swarm network out = swarmnet.list_agents() print(out) # Run the workflow on a task out = swarmnet.run_single_agent( agent2.id, "Generate a 10,000 word blog on health and wellness." ) print(out) # Run all the agents in the swarm network on a task out = swarmnet.run_many_agents("Generate a 10,000 word blog on health and wellness.") print(out) ``` ### Majority Voting Multiple-agents will evaluate an idea based off of an parsing or evaluation function. From papers like "[More agents is all you need](https://arxiv.org/pdf/2402.05120.pdf) ```python from swarms import Agent, MajorityVoting, ChromaDB, Anthropic # Initialize the llm llm = Anthropic() # Agents agent1 = Agent( llm = llm, system_prompt="You are the leader of the Progressive Party. What is your stance on healthcare?", agent_name="Progressive Leader", agent_description="Leader of the Progressive Party", long_term_memory=ChromaDB(), max_steps=1, ) agent2 = Agent( llm=llm, agent_name="Conservative Leader", agent_description="Leader of the Conservative Party", long_term_memory=ChromaDB(), max_steps=1, ) agent3 = Agent( llm=llm, agent_name="Libertarian Leader", agent_description="Leader of the Libertarian Party", long_term_memory=ChromaDB(), max_steps=1, ) # Initialize the majority voting mv = MajorityVoting( agents=[agent1, agent2, agent3], output_parser=llm.majority_voting, autosave=False, verbose=True, ) # Start the majority voting mv.run("What is your stance on healthcare?") ``` ### Real-World Deployment ### Multi-Agent Swarm for Logistics Here's a production grade swarm ready for real-world deployment in a factory and logistics settings like warehouses. This swarm can automate 3 costly and inefficient workflows, safety checks, productivity checks, and warehouse security. ```python import os from dotenv import load_dotenv from swarms.models import GPT4VisionAPI from swarms.prompts.logistics import ( Efficiency_Agent_Prompt, Health_Security_Agent_Prompt, Productivity_Agent_Prompt, Quality_Control_Agent_Prompt, Safety_Agent_Prompt, Security_Agent_Prompt, Sustainability_Agent_Prompt, ) from swarms.structs import Agent # Load ENV load_dotenv() api_key = os.getenv("OPENAI_API_KEY") # GPT4VisionAPI llm = GPT4VisionAPI(openai_api_key=api_key) # Image for analysis factory_image = "factory_image1.jpg" # Initialize agents with respective prompts health_security_agent = Agent( llm=llm, sop=Health_Security_Agent_Prompt, max_loops=1, multi_modal=True, ) # Quality control agent quality_control_agent = Agent( llm=llm, sop=Quality_Control_Agent_Prompt, max_loops=1, multi_modal=True, ) # Productivity Agent productivity_agent = Agent( llm=llm, sop=Productivity_Agent_Prompt, max_loops=1, multi_modal=True, ) # Initiailize safety agent safety_agent = Agent(llm=llm, sop=Safety_Agent_Prompt, max_loops=1, multi_modal=True) # Init the security agent security_agent = Agent( llm=llm, sop=Security_Agent_Prompt, max_loops=1, multi_modal=True ) # Initialize sustainability agent sustainability_agent = Agent( llm=llm, sop=Sustainability_Agent_Prompt, max_loops=1, multi_modal=True, ) # Initialize efficincy agent efficiency_agent = Agent( llm=llm, sop=Efficiency_Agent_Prompt, max_loops=1, multi_modal=True, ) # Run agents with respective tasks on the same image health_analysis = health_security_agent.run( "Analyze the safety of this factory", factory_image ) quality_analysis = quality_control_agent.run( "Examine product quality in the factory", factory_image ) productivity_analysis = productivity_agent.run( "Evaluate factory productivity", factory_image ) safety_analysis = safety_agent.run( "Inspect the factory's adherence to safety standards", factory_image, ) security_analysis = security_agent.run( "Assess the factory's security measures and systems", factory_image, ) sustainability_analysis = sustainability_agent.run( "Examine the factory's sustainability practices", factory_image ) efficiency_analysis = efficiency_agent.run( "Analyze the efficiency of the factory's manufacturing process", factory_image, ) ``` --- ## Build your own LLMs, Agents, and Swarms! ### Swarms Compliant Model Interface ```python from swarms import BaseLLM class vLLMLM(BaseLLM): def __init__(self, model_name='default_model', tensor_parallel_size=1, *args, **kwargs): super().__init__(*args, **kwargs) self.model_name = model_name self.tensor_parallel_size = tensor_parallel_size # Add any additional initialization here def run(self, task: str): pass # Example model = vLLMLM("mistral") # Run the model out = model("Analyze these financial documents and summarize of them") print(out) ``` ### Swarms Compliant Agent Interface ```python from swarms import Agent class MyCustomAgent(Agent):     def __init__(self, *args, **kwargs):         super().__init__(*args, **kwargs)         # Custom initialization logic     def custom_method(self, *args, **kwargs):         # Implement custom logic here         pass     def run(self, task, *args, **kwargs):         # Customize the run method         response = super().run(task, *args, **kwargs)         # Additional custom logic         return response` # Model agent = MyCustomAgent() # Run the agent out = agent("Analyze and summarize these financial documents: ") print(out) ``` ### Compliant Interface for Multi-Agent Collaboration ```python from swarms import AutoSwarm, AutoSwarmRouter, BaseSwarm # Build your own Swarm class MySwarm(BaseSwarm): def __init__(self, name="kyegomez/myswarm", *args, **kwargs): super().__init__(*args, **kwargs) self.name = name def run(self, task: str, *args, **kwargs): # Add your multi-agent logic here # agent 1 # agent 2 # agent 3 return "output of the swarm" # Add your custom swarm to the AutoSwarmRouter router = AutoSwarmRouter( swarms=[MySwarm] ) # Create an AutoSwarm instance autoswarm = AutoSwarm( name="kyegomez/myswarm", description="A simple API to build and run swarms", verbose=True, router=router, ) # Run the AutoSwarm autoswarm.run("Analyze these financial data and give me a summary") ``` ## `AgentRearrange` Inspired by Einops and einsum, this orchestration techniques enables you to map out the relationships between various agents. For example you specify linear and sequential relationships like `a -> a1 -> a2 -> a3` or concurrent relationships where the first agent will send a message to 3 agents all at once: `a -> a1, a2, a3`. You can customize your workflow to mix sequential and concurrent relationships. [Docs Available:](https://swarms.apac.ai/en/latest/swarms/structs/agent_rearrange/) ```python from swarms import Agent, AgentRearrange, rearrange, 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="", 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="", 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="", 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) # Using rearrange function output = rearrange( agents, flow, "Create a format to express and communicate swarms of llms in a structured manner for youtube", ) print(output) ``` ## `HierarhicalSwarm` Coming soon... ## `AgentLoadBalancer` Coming soon... ## `GraphSwarm` Coming soon... ## `MixtureOfAgents` This is an implementation from the paper: "Mixture-of-Agents Enhances Large Language Model Capabilities" by together.ai, it achieves SOTA on AlpacaEval 2.0, MT-Bench and FLASK, surpassing GPT-4 Omni. Great for tasks that need to be parallelized and then sequentially fed into another loop ```python 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="", 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="", 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="", 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) ``` --- ## Documentation Documentation is located here at: [swarms.apac.ai](https://swarms.apac.ai) ---- ## 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`. ```sh ├── __init__.py ├── agents ├── artifacts ├── memory ├── schemas ├── 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](/CONTRIBUTING.md). Bug Report? [File here](https://github.com/swarms/gateway/issues) | Feature Request? [File here](https://github.com/swarms/gateway/issues) 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](https://github.com/kyegomez/swarms/blob/master/CONTRIBUTING.md) and our [contributing board](https://github.com/users/kyegomez/projects/1) to participate in Roadmap discussions! ---- ## Community Join our growing community around the world, for real-time support, ideas, and discussions on Swarms 😊 - View our official [Blog](https://swarms.apac.ai) - Chat live with us on [Discord](https://discord.gg/kS3rwKs3ZC) - Follow us on [Twitter](https://twitter.com/kyegomez) - Connect with us on [LinkedIn](https://www.linkedin.com/company/the-swarm-corporation) - Visit us on [YouTube](https://www.youtube.com/channel/UC9yXyitkbU_WSy7bd_41SqQ) - [Join the Swarms community on Discord!](https://discord.gg/AJazBmhKnr) - 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](https://lu.ma/5p2jnc2v) --- ## Discovery Call Book a discovery call to learn how Swarms can lower your operating costs by 40% with swarms of autonomous agents in lightspeed. [Click here to book a time that works for you!](https://calendly.com/swarm-corp/30min?month=2023-11) ## Accelerate Backlog Accelerate Bugs, Features, and Demos to implement by supporting us here: ## Docker Instructions - [Learn More Here About Deployments In Docker](https://swarms.apac.ai/en/latest/docker_setup/) ## Swarm Newsletter 🤖 🤖 🤖 📧 Sign up to the Swarm newsletter to receive updates on the latest Autonomous agent research papers, step by step guides on creating multi-agent app, and much more Swarmie goodiness 😊 [CLICK HERE TO SIGNUP](https://docs.google.com/forms/d/e/1FAIpQLSfqxI2ktPR9jkcIwzvHL0VY6tEIuVPd-P2fOWKnd6skT9j1EQ/viewform?usp=sf_link) # License Apache License # Citations Please cite Swarms in your paper or your project if you found it beneficial in any way! Appreciate you. ```bibtex @misc{swarms, author = {Gomez, Kye}, title = {{Swarms: The Multi-Agent Collaboration Framework}}, howpublished = {\url{https://github.com/kyegomez/swarms}}, year = {2023}, note = {Accessed: Date} } ``` ```bibtex @misc{wang2024mixtureofagents, title={Mixture-of-Agents Enhances Large Language Model Capabilities}, author={Junlin Wang and Jue Wang and Ben Athiwaratkun and Ce Zhang and James Zou}, year={2024}, eprint={2406.04692}, archivePrefix={arXiv}, primaryClass={cs.CL} } ```