Autonomous AI Coding Agent Uses MCTS and Gemini for Secure Code

The evolving demands of software engineering highlight a critical need for automated tools capable of generating secure source code from natural language descriptions. Traditional Large Language Models (LLMs) often fall short in complex operations due to their "one-shot" generation limitations, leading to logical errors and reduced algorithmic efficiency. This research introduces an autonomous AI Coding Agent designed to bridge the gap between LLM output and production-grade software. The proposed framework leverages the Gemini 2.5 Flash API for its core reasoning capabilities. To overcome the limitations of direct LLM generation, it employs a customized Monte Carlo Tree Search (MCTS) method, treating code generation as a sophisticated search problem. This allows the agent to explore multiple implementation paths and refine its solutions. A "Self-Critic" evaluator system is integrated to test and rank different code implementations based on accuracy and difficulty. The agent then uses backpropagation to improve its operational framework. Operating via a Flask-based web interface, the system provides instant feedback and syntax highlighting. Experimental results demonstrate a 92% success rate on complex logical prompts, significantly surpassing the performance of standard zero-shot generation models.