🤖 AI-Driven Self-Evolving Software

📖 Introduction

This repository contains the implementation and experimental code for the paper:

AI-Driven Self-Evolving Software

This repository provides:

• ✅ A working implementation of the core algorithm presented in the paper
• ⚙️ Scripts and configuration files to reproduce the reported experiments

---

📝 Abstract

Software automation has long been a central goal of software engineering, aiming to minimize human intervention in the development lifecycle. Recent efforts have leveraged Artificial Intelligence (AI) to advance software automation with notable progress. However, current AI mainly serves as an assistant to human developers, leaving software development still dependent on explicit human input.

This raises a fundamental question:

Can AI move beyond its role as an assistant to become a core component of software, enabling genuine software automation?

To investigate this, we introduce AI-Driven Self-Evolving Software — a new form of software that evolves continuously through direct interaction with users.

We demonstrate this vision with a lightweight prototype built on a multi-agent architecture that:

• 🧩 Interprets user requirements autonomously
• 🖥️ Generates and validates code
• 🔄 Integrates new functionalities seamlessly

Case studies across multiple scenarios show that the prototype can reliably construct and reuse functionality, providing early evidence that such systems can scale to sophisticated applications — paving the way toward truly automated software development.

---

🎬 Demo

---

⚡ Requirements

• Python ≥ 3.12.11
• Install dependencies:

pip install -r requirements.txt

---

📂 Directory Layout

repo/
├── src-v2.0/        # Source code for cross validation
├── src/        # Source code for unit-test
├── config.json # Configuration file
├── requirements.txt
├── demo.gif    # Demo GIF
├── figures
├── output.log # more cases provided in paper
└── README.md

---

🚀 Quick Start

1. Edit the api_key and base_url in the config.json file
2. Run the agent system:

python -m src.core.agent

---

✅ Tip: Make sure your Python environment is correctly set up before running the agent.

📋 Cases

Using weather prediction as an example, we explain how our method operates in detail. Additional examples can be found in the output.log file.

Initially, when no user operations have been performed, the system contains only a Leader node. The Data Manager maintains the Meta Node at this stage.

Code structure: {
    "type": "dir", 
    "children": {}
}

In the first round of interaction, the user submits an initial query: "Please help me check the weather in Beijing for tomorrow and the day after tomorrow."

Upon receiving the user's request, the Leader node intercepts both the query and the current data structure from the Data Manager. It then analyzes this information to generate subsequent instructions.

Based on this analysis, the Leader then dispatches the task to the Code Generator.

Once the Leader designates the Code Generator as the next step, the Code Generator produces the requisite code. This generated code is subsequently returned to the Leader, which then delegates it to the Code Validator for testing.

Upon completion of cross-validation among multiple code candidates and testing programs, the code with the highest confidence score is selected. This optimal code is subsequently persisted to the Data Manager, triggering a structural update.

Code structure: {
    'type': 'dir', 
    'children': {
        'weather_forecast.py': {
            'type': 'file', 
            'classes': [
                {
                'name': '_HTTPClient', 
                'members': ['_timeout'], 
                'methods': ['__init__', 'get_json']
                }
            ], 
            'functions': [
                '_geocode', 
                '_build_forecast_params', 
                'fetch_weather', 
                '_parse_cli', 
                'main'
            ], 
            'dependencies': [
                'requests', 
                'argparse', 
                'datetime', 
                'json', 
                'sys', 
                'typing'
            ], 
            'description': 'weather_forecast.py is a self-contained module exposing fetch_weather() and a CLI\nentry-point. It hides network details behind _HTTPClient, uses Open-Meteo APIs\nfor geocoding and daily forecasts, maps weather codes to human-readable\nconditions, and returns/prints a compact JSON string. The module validates\narguments, handles HTTP errors, and limits requests to 7 days ahead. Usage:\n    python weather_forecast.py --city "Beijing" --days 2\nOr programmatically:\n    from weather_forecast import fetch_weather\n    data_json = fetch_weather("Beijing", 2)'
        }
    }
}

The Code Runner ultimately executes the code to generate the response for the user.

---

In the second round of interaction, the user submits an query: "I am currently in London, please help me check the weather for the next two days in London."

As the structure maintained by the Data Manager has been updated following the previous query, the Leader now determines that the Code Generator can be bypassed, and the task can be directed straight to the Code Runner for execution.