Concurrent Generative-Analytic Swarms: A Hierarchical Agent-Based Approach for Efficient Multi-Modal Synthesis and Assessment

The implementation of: ``Harnessing Hierarchical Swarms: A Groundbreaking Approach to Multi-Modal Content Creation''. https://medium.com/@arash.mansoori65/harnessing-hierarchical-swarms-a-groundbreaking-approach-to-multi-modal-content-creation-061dbcb698f6

Description

The proposed method introduces a novel framework for a hierarchical swarm of agents designed to tackle composite tasks with high concurrency and efficiency. The framework comprises three specialized agents: the Controller Agent, the Generator Agent, and the Analyzer Agent, each fulfilling distinctive roles within task decomposition, content generation, and content analysis.

Agent Interaction Model

The Controller Agent lies at the apex of the hierarchy and is responsible for decomposing composite tasks into two parallel subtasks. One subtask targets content generation, managed by the Generator Agent, while the other concerns content analysis, overseen by the Analyzer Agent. The diagram below illustrates the agent interactions according to the established workflow.

sequenceDiagram
    participant CA as Controller Agent
    participant GA as Generator Agent
    participant AA as Analyzer Agent
    
    CA->>GA: Assign Generation Subtask
    CA->>AA: Assign Analysis Subtask
    
    par
        GA->>GA: Concurrent Content Generation
        GA-->>AA: Send Generated Content
        AA->>AA: Concurrent Content Analysis
        AA-->>CA: Send Analysis Results
    end

Concurrent Generation and Analysis Process

This section expounds on the concurrent content generation and analysis processes to highlight their saliency in achieving high throughput and real-time task execution.

Concurrent Content Generation

Following task assignment, the Generator Agent simultaneously generates multiple content instances. This not only streamlines the content creation process but also serves to enhance overall system productivity.

flowchart TB
    CA[Controller Agent] -->|Assigns subtasks| GA[Generator Agent]
    GA -->|Generates content concurrently| G1[Generation Instance 1]
    GA -->|Generates content concurrently| G2[Generation Instance 2]
    GA -->|Generates content concurrently| GN[Generation Instance N]

Concurrent Content Analysis

Similarly, the Analyzer Agent conducts concurrent analysis of content generated by the Generator Agent leveraging the following communication model:

flowchart TB
    GA[Generator Agent] -- Generated Content --> AA[Analyzer Agent]
    AA --> A1[Analysis Instance 1]
    AA --> A2[Analysis Instance 2]
    AA --> AN[Analysis Instance N]
    A1 -->|Assessment| CA[Controller Agent]
    A2 -->|Assessment| CA
    AN -->|Assessment| CA

The Analyzer Agent processes each generation concurrently, acknowledges receipt to the Generator Agent, and communicates the assessment outcomes to the Controller Agent.

Future Extension Model

Consideration is given to future workflow enhancements to increase robustness and scalability. The integration of a central database allows for decoupling of content generation from analysis, potentially reducing bottlenecks and accommodating asynchronous operations. The current code base save generation and analysis outputs locally. The integration of database inclusion into the process is straightforward and is deferred to future developments.

sequenceDiagram
    participant CA as Controller Agent
    participant GA as Generator Agent
    participant DB as Database
    participant AA as Analyzer Agent
    
    CA->>GA: Assign Generation Subtask
    GA->>GA: Concurrent Content Generation
    GA->>DB: Store Generated Content
    AA->>DB: Pull Data for Analysis
    AA->>AA: Concurrent Content Analysis
    AA->>CA: Send Analysis Results

Note: In the described workflow, it's important to note that the processes of generation and analysis operate independently. Consequently, there is no requisite for a controller agent since the commands for the generator and analyzer agents can be issued distinctly.

Algorithmic Description of the Proposed Method

The algorithmic process is outlined as follows, with a clear connection to the previously introduced block diagrams for clarity and ease of understanding:

Task Decomposition:
- The Controller Agent receives a composite task.
- The composite task is decomposed into a content generation subtask and a content analysis subtask.
Concurrent Content Generation:
- The Generator Agent conducts content generation for the subtask concurrently across multiple instances, N.
Concurrent Content Analysis:
- The Analyzer Agent, having received content concurrently, analyzes them in parallel across its multiple analysis instances, N.
Communication Protocol:
- Upon completion of generation, the Generator Agent sends the content to the Analyzer Agent.
- After analysis, the Analyzer Agent sends an acknowledgment back to the Generator Agent and forwards the analysis results to the Controller Agent.
Future Extension Workflow:
- The system can be extended to integrate a database where generated content is stored.
- The Analyzer Agent pulls content from this database enabling asynchronous analysis, which may facilitate scalability and flexibility.

Installation

Install all the necessary requirements.

pip install -r requirements.txt

Environment Variables

Create a .env file and set your openai API key.

Run the Swarm of Agents

Set the appropriate parameters according to your need in settings.py. Then, run the multi-modal swarm of agents.

python run_agents_swarm.py

For asynchronous generation and analysis, run the following.

python run_async_gen_anlys.py

Finally, run the following for the baseline generation and analysis, respectively.

python run_gen_baseline.py
python run_anlys_baseline.py

To use the docker compose for running the swarm of agents, use the following command. Note that you need to use the corresponding functions for using docker, these functions are the ones ending with _docker in this repository.

docker compose up -d

To follow log output(s) from containers use the following command.

docker compose logs -f

To stop and remove container(s), network(s), volume(s), and image(s) created by up, use the following command.

docker compose down -v

Remove Registered Agents

To remove the registered agents after running the swarm of agents, run the following command.

python run_remove_agents.py

Results

The results for this research are saved in data and data_async with the subdirectories generation and analysis. The saved files are self explanatory.

Author

Arash Shahmansoori (arash.mansoori65@gmail.com)

License

This project is licensed under the terms of the MIT License

arash-shahmansoori/multi_modal_agents_swarm