MixtureOfAgents Class Documentation¶

Overview¶

The MixtureOfAgents class represents a mixture of agents operating within a swarm. The workflow of the swarm follows a parallel → sequential → parallel → final output agent process. This implementation is inspired by concepts discussed in the paper: https://arxiv.org/pdf/2406.04692.

The class is designed to manage a collection of agents, orchestrate their execution in layers, and handle the final aggregation of their outputs through a designated final agent. This architecture facilitates complex, multi-step processing where intermediate results are refined through successive layers of agent interactions.

Class Definition¶

MixtureOfAgents¶

class MixtureOfAgents(BaseSwarm):

Attributes¶

Attribute	Type	Description	Default
`agents`	`List[Agent]`	The list of agents in the swarm.	`None`
`flow`	`str`	The flow of the swarm.	`parallel -> sequential -> parallel -> final output agent`
`max_loops`	`int`	The maximum number of loops to run.	`1`
`verbose`	`bool`	Flag indicating whether to print verbose output.	`True`
`layers`	`int`	The number of layers in the swarm.	`3`
`rules`	`str`	The rules for the swarm.	`None`
`final_agent`	`Agent`	The agent to handle the final output processing.	`None`
`auto_save`	`bool`	Flag indicating whether to auto-save the metadata to a file.	`False`
`saved_file_name`	`str`	The name of the file where the metadata will be saved.	`"moe_swarm.json"`

Methods¶

`init`¶

Parameters¶

Parameter	Type	Description	Default
`name`	`str`	The name of the swarm.	`"MixtureOfAgents"`
`description`	`str`	A brief description of the swarm.	`"A swarm of agents that run in parallel and sequentially."`
`agents`	`List[Agent]`	The list of agents in the swarm.	`None`
`max_loops`	`int`	The maximum number of loops to run.	`1`
`verbose`	`bool`	Flag indicating whether to print verbose output.	`True`
`layers`	`int`	The number of layers in the swarm.	`3`
`rules`	`str`	The rules for the swarm.	`None`
`final_agent`	`Agent`	The agent to handle the final output processing.	`None`
`auto_save`	`bool`	Flag indicating whether to auto-save the metadata to a file.	`False`
`saved_file_name`	`str`	The name of the file where the metadata will be saved.	`"moe_swarm.json"`

`agent_check`¶

def agent_check(self):

Description¶

Checks if the provided agents attribute is a list of Agent instances. Raises a TypeError if the validation fails.

Example Usage¶

moe_swarm = MixtureOfAgents(agents=[agent1, agent2])
moe_swarm.agent_check()  # Validates the agents

`final_agent_check`¶

def final_agent_check(self):

Description¶

Checks if the provided final_agent attribute is an instance of Agent. Raises a TypeError if the validation fails.

Example Usage¶

moe_swarm = MixtureOfAgents(final_agent=final_agent)
moe_swarm.final_agent_check()  # Validates the final agent

`swarm_initialization`¶

def swarm_initialization(self):

Description¶

Initializes the swarm by logging the swarm name, description, and the number of agents.

Example Usage¶

moe_swarm = MixtureOfAgents(agents=[agent1, agent2])
moe_swarm.swarm_initialization()  # Initializes the swarm

`run`¶

def run(self, task: str = None, *args, **kwargs):

Parameters¶

Parameter	Type	Description	Default
`task`	`str`	The task to be performed by the swarm.	`None`
`*args`	`tuple`	Additional arguments.	`None`
`**kwargs`	`dict`	Additional keyword arguments.	`None`

Returns¶

Type	Description
`str`	The conversation history as a string.

Description¶

Runs the swarm with the given task, orchestrates the execution of agents through the specified layers, and returns the conversation history.

Example Usage¶

moe_swarm = MixtureOfAgents(agents=[agent1, agent2], final_agent=final_agent)
history = moe_swarm.run(task="Solve this problem.")
print(history)

Detailed Explanation¶

Initialization¶

The __init__ method initializes the swarm with the provided parameters, sets up the conversation rules, and invokes the initialization of the swarm. It also ensures the validity of the agents and final_agent attributes by calling the agent_check and final_agent_check methods respectively.

Agent Validation¶

The agent_check method validates whether the agents attribute is a list of Agent instances, while the final_agent_check method validates whether the final_agent is an instance of Agent. These checks are crucial to ensure that the swarm operates correctly with the appropriate agent types.

Swarm Initialization¶

The swarm_initialization method logs essential information about the swarm, including its name, description, and the number of agents. This provides a clear starting point for the swarm's operations and facilitates debugging and monitoring.

Running the Swarm¶

The run method is the core of the MixtureOfAgents class. It orchestrates the execution of agents through multiple layers, collects their outputs, and processes the final output using the final_agent. The conversation history is maintained and updated throughout this process, allowing for a seamless flow of information and responses.

During each layer, the method iterates over the agents, invokes their run method with the current conversation history, and logs the outputs. These outputs are then added to the conversation, and the history is updated for the next layer.

After all layers are completed, the final output agent processes the entire conversation history, and the metadata is created and optionally saved to a file. This metadata includes details about the layers, agent runs, and final output, providing a comprehensive record of the swarm's execution.

Additional Information and Tips¶

Common Issues and Solutions¶

Type Errors: Ensure that all agents in the agents list and the final_agent are instances of the Agent class. The agent_check and final_agent_check methods help validate this.
Verbose Logging: Use the verbose flag to control the verbosity of the output. This can help with debugging or reduce clutter in the logs.
Auto-Save Feature: Utilize the auto_save flag to automatically save the metadata to a file. This can be useful for keeping records of the swarm's operations without manual intervention.

References and Resources¶

For further reading and background information on the concepts used in the MixtureOfAgents class, refer to the paper: https://arxiv.org/pdf/2406.04692.

Usage Examples¶

Example 1: Basic Initialization and Run¶

from swarms import MixtureOfAgents, Agent

from swarm_models import OpenAIChat

# Define agents
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",
)


# Initialize the MixtureOfAgents
moe_swarm = MixtureOfAgents(agents=[director, accountant1, accountant2], final_agent=director)

# Run the swarm
history = moe_swarm.run(task="Perform task X.")
print(history)

Example 2: Verbose Output and Auto-Save¶

from swarms import MixtureOfAgents, Agent

from swarm_models import OpenAIChat

# Define Agents
# Define agents
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",
)

# Initialize the MixtureOfAgents with verbose output and auto-save enabled
moe_swarm = MixtureOfAgents(
    agents=[director, accountant1, accountant2],
    final_agent=director,
    verbose=True,
    auto_save=True
)

# Run the swarm
history = moe_swarm.run(task="Analyze data set Y.")
print(history)

Example 3: Custom Rules and Multiple Layers¶

from swarms import MixtureOfAgents, Agent

from swarm_models import OpenAIChat

# Define agents
# 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",
)

# Initialize the MixtureOfAgents with custom rules and multiple layers
moe_swarm = MixtureOfAgents(
    agents=[director, accountant1, accountant2],
    final_agent=director,
    layers=5,
    rules="Custom rules for the swarm"
)

# Run the swarm
history = moe_swarm.run(task="Optimize process Z.")
print(history)

This comprehensive documentation provides a detailed understanding of the MixtureOfAgents class, its attributes, methods, and usage. The examples illustrate how to initialize and run the swarm, demonstrating its flexibility and capability to handle various tasks and configurations.

Conclusion¶

The MixtureOfAgents class is a powerful and flexible framework for managing and orchestrating a swarm of agents. By following a structured approach of parallel and sequential processing, it enables the implementation of complex multi-step workflows where intermediate results are refined through multiple layers of agent interactions. This architecture is particularly suitable for tasks that require iterative processing, collaboration among diverse agents, and sophisticated aggregation of outputs.

Key Takeaways¶

Flexible Initialization: The class allows for customizable initialization with various parameters, enabling users to tailor the swarm's configuration to their specific needs.
Robust Agent Management: With built-in validation methods, the class ensures that all agents and the final agent are correctly instantiated, preventing runtime errors and facilitating smooth execution.
Layered Processing: The layered approach to processing allows for intermediate results to be iteratively refined, enhancing the overall output quality.
Verbose Logging and Auto-Save: These features aid in debugging, monitoring, and record-keeping, providing transparency and ease of management.
Comprehensive Documentation: The detailed class and method documentation, along with numerous usage examples, provide a clear and thorough understanding of how to leverage the MixtureOfAgents class effectively.

Practical Applications¶

The MixtureOfAgents class can be applied in various domains, including but not limited to:

Natural Language Processing (NLP): Managing a swarm of NLP models to process, analyze, and synthesize text.
Data Analysis: Coordinating multiple data analysis agents to process and interpret complex data sets.
Optimization Problems: Running a swarm of optimization algorithms to solve complex problems in fields such as logistics, finance, and engineering.
AI Research: Implementing experimental setups that require the collaboration of multiple AI models or agents to explore new methodologies and approaches.

Future Extensions¶

The MixtureOfAgents framework provides a solid foundation for further extensions and customizations, including:

Dynamic Layer Configuration: Allowing layers to be added or removed dynamically based on the task requirements or intermediate results.
Advanced Agent Communication: Enhancing the communication protocols between agents to allow for more sophisticated information exchange.
Integration with Other Frameworks: Seamlessly integrating with other machine learning or data processing frameworks to leverage their capabilities within the swarm architecture.

In conclusion, the MixtureOfAgents class represents a versatile and efficient solution for orchestrating multi-agent systems, facilitating complex task execution through its structured and layered approach. By harnessing the power of parallel and sequential processing, it opens up new possibilities for tackling intricate problems across various domains.

MixtureOfAgents Class Documentation¶

Overview¶

Class Definition¶

MixtureOfAgents¶

Attributes¶

Methods¶

__init__¶

Parameters¶

agent_check¶

Description¶

Example Usage¶

final_agent_check¶

Description¶

Example Usage¶

swarm_initialization¶

Description¶

Example Usage¶

run¶

Parameters¶

Returns¶

Description¶

Example Usage¶

Detailed Explanation¶

Initialization¶

Agent Validation¶

Swarm Initialization¶

Running the Swarm¶

Additional Information and Tips¶

Common Issues and Solutions¶

References and Resources¶

Usage Examples¶

Example 1: Basic Initialization and Run¶

Example 2: Verbose Output and Auto-Save¶

Example 3: Custom Rules and Multiple Layers¶

Conclusion¶

Key Takeaways¶

Practical Applications¶

Future Extensions¶

`init`¶

`agent_check`¶

`final_agent_check`¶

`swarm_initialization`¶

`run`¶