Autonomous AI Software Development Team (AiDevSquad)

This project implements a multi-agent AI system designed to autonomously develop, review, test, and document software projects from a high-level user prompt. It simulates a team of AI agents, each with a specialized role, working together to deliver a complete, version-controlled software application.

🚀 Features

• Multi-Agent Architecture: Simulates a real development team with specialized agents:

  • SADTSARTPlannerAgent: The primary planner that uses SADT/SART methodology to generate hierarchical workplans and atomic actions.
  • ProductOwnerAgent: Acts as the vision holder and requirements consultant, answering agent questions to clarify ambiguity.
  • DeveloperAgent: Executes tasks using a ReAct (Reason+Act) model and a suite of tools.
  • CodeReviewerAgent: Analyzes the entire codebase for integration issues, bugs, and inconsistencies.
  • UnitTestAgent: Generates pytest unit tests for the generated code.
  • DocumentationAgent: Writes the final project README.md file.

• Flexible Agent Collaboration: Agents can now communicate dynamically using the call_agent tool. For example, a DeveloperAgent can ask the CodeReviewerAgent for immediate feedback on a specific file, breaking the rigid sequential handoff structure.

• Tool-Based Actions (ReAct): The DeveloperAgent uses a set of tools (list_files, read_file, write_file, call_agent, etc.) to interact with the codebase and other agents.

• Automated Version Control: Every completed task is automatically committed to a Git repository, providing a complete, auditable history of the AI's work.

• RAG-Powered Context: Uses Retrieval-Augmented Generation (RAG) via LlamaIndex to provide agents with relevant code context, overcoming LLM context window limitations.

• Pluggable LLM Providers: Easily switch between local models (via Ollama) and cloud-based APIs (like Google Gemini).

• Cost & Usage Tracking: Monitors token usage and estimates costs for each LLM call.

• Structured Trace Logging: Generates detailed Markdown and JSON execution traces for easy debugging and monitoring.

🏛️ System Architecture

The system is orchestrated by a central Orchestrator class that manages the project lifecycle through several distinct phases:

1. Planning: The SADTSARTPlannerAgent analyzes the requirement and generates a detailed hierarchical plan (SADT/SART). The ProductOwnerAgent is available for consultation throughout the project.
2. Development & Collaboration: The Orchestrator iterates through each task in the plan. The DeveloperAgent executes the task using its tools. Crucially, the DeveloperAgent is responsible for quality assurance. It can use the call_agent tool to request ad-hoc reviews or advice from the CodeReviewerAgent or ProductOwnerAgent at any time during the task.
3. Unit Testing (Optional): If enabled, the UnitTestAgent generates pytest files for the Python code.
4. Final Validation: The TesterAgent runs the test suite (if any) and attempts to start the main application to ensure it's executable.
5. Documentation: If all previous phases succeed, the DocumentationAgent generates a README.md file for the newly created project.

🛠️ Setup & Installation

Follow these steps to set up and run the project.

Prerequisites

• Python 3.11+
• Git
• Docker (for the TesterAgent)
• (Optional) Ollama for running local models.

Installation

1. Clone the repository:

   git clone https://github.com/sancelot/AIdevSquad
   cd AIdevSquad

2. Create a virtual environment and install dependencies:

   python -m venv .venv
   source .venv/bin/activate  # On Windows, use: .venv\Scripts\activate
   pip install -r requirements.txt

3. Set up LLM providers:

   • Create a .env file in the root directory by copying the example:

     cp .env.example .env

   • For Google Gemini:
     • Go to Google AI Studio to get an API key.
     • Add it to your .env file: GOOGLE_API_KEY="your-api-key"
   • For Ollama (Local Models):
     • Install Ollama.
     • Pull the models you want to use. We recommend:

       ollama pull llama3:8b
       ollama pull bge-base-en-v1.5  # For embeddings

▶️ How to Run

The main script is orchestrator.py. You can control its behavior using command-line flags.

Basic Run (using the default provider, e.g., Ollama):

python orchestrator.py myproject --prompt myrequirements.txt

Run with Ollama:

python orchestrator.py myproject --ollama --prompt myrequirements.txt

Start a new project from scratch (deletes the existing workspace):

python orchestrator.py myproject --new --prompt myrequirements.txt

**Generate unit tests as part of the process:**

```bash
python orchestrator.py myproject--with-tests --prompt myrequirements.txt

Flags can be combined:

python orchestrator.py myproject --ollama --new --with-tests --prompt myrequirements.txt

After each run, a detailed trace log is saved in the logs/ directory, and the generated project (with its own Git repository) is available in the ..._workspace/ directory.

Monitor the Results with the UI

A simple Flask-based UI is available to visualize the execution traces.

1. Navigate to the UI directory:

   cd monitoring_ui

2. Install its dependencies (if any):

   pip install -r requirements.txt

3. Run the UI application:

   flask run --port=5001

4. Open your browser to http://127.0.0.1:5001 to see the dashboard.

New Feature: Nested Collaboration View The trace view now supports visualizing nested agent calls. When a DeveloperAgent consults a CodeReviewerAgent or ProductOwnerAgent via chat, the sub-agent's thoughts and tool usage are displayed with indentation and distinct styling, providing a clear "call tree" of the collaboration.

🤝 Contributing

Contributions are welcome! Please feel free to open an issue or submit a pull request. Areas for improvement include:

• Providing consistence documentation between agents in order to avoid producing a project that finally builds but is not usable.
• Adding more tools for the DeveloperAgent (e.g., AST-based refactoring for specific languages).
• Improving the intelligence of the CodeReviewerAgent.
• Expanding the TesterAgent to run more complex integration tests.
• Enhancing the monitoring UI.

📜 License

This project is licensed under the MIT License. See the LICENSE file for details.