Proposal for AI-Enhanced Multi-Agent System for Automated Upstream Contributions to MariaDB/Server

Presentation materials

Abstract

This proposal presents an innovative three-phase multi-agent system integrated with Amazon’s Model Context Protocol (MCP) to automate contributions to the MariaDB/Server open-source project. In the first phase, a single-agent AI agent, the Upstream Pilot, to rank, analyze and suggest MariaDB/Server Jira tickets. The innovative approach performs a bulk rank based on Jira tickets labels and adds analysis and implementation suggestions to the user through a Tampermonkey script.

In the second phase, the system supports developers by automating repetitive and context-intensive tasks such as retrieval, planning, coding, debugging, and code reviews, significantly enhancing developer productivity while ensuring human oversight for complex decisions.

In the third phase, the system evolves into a largely autonomous solution by employing advanced AI methodologies, including reinforcement learning, meta-learning, and human-in-the-loop fine-tuning. These methodologies allow the system to dynamically adapt its planning, execution, and debugging strategies based on continuous learning from historical data and structured developer feedback, progressively reducing the need for manual interventions.

By leveraging sophisticated AWS infrastructure components like Amazon Bedrock, AWS Lambda, AWS Step Functions, and Amazon OpenSearch Serverless, the proposed multi-agent system achieves robust scalability and adaptability. The implementation of advanced checkpoint debugging techniques further enhances accuracy and reduces error rates. Ultimately, this system aims to dramatically increase Amazon’s efficiency and quality of contributions to the MariaDB open-source community, fostering a more collaborative and productive development environment.

Speaker: Bardia Hassanzadeh, Software Development Engineer @Amazon Web Services (AWS)