Multi-Agent Code Generation System
Project PREP0005296 · NIST sponsor Shengyen Li
Overview
The current AI system is cost-effective for prototyping software code, but code reliability and adoption still depend on the users’ benchmark dataset and the prompting instance. The proposed system introduces a multi-agent AI architecture for code generation, providing scalable, robust capabilities to manage and analyze datasets via tools such as the Model Context Protocol (MCP), Claude Code, and AutoGen. To ensure a competitive advantage, we will conduct a comprehensive GitHub survey and risk assessment of similar tools currently available. The success of this project will provide a benchmark example for automated code generation, validation, and create a smart AI system for additive manufacturing.
Qualifications
- An undergraduate student currently enrolled in Computer Science, Engineering, Manufacturing, or a related field.
- Familiarity with coding environment, including Python, Command Line Tools, and AI platforms.
- Familiarity with methods for statistical analysis.
- Ability to develop prototypes of tools needed to analyze data.
- Strong oral and written communication skills.
Research Proposal
Key responsibilities will include but are not limited to:
- Create a generator-Validator Pair: The system uses a multi-agent architecture in which one AI agent generates code and a second, independent agent validates it, etc.
- Develop an iterative feedback loop: Instead of a single-pass execution, the agents operate in a continuous cycle. If the Validator detects an error, it returns the execution logs to the Generator for targeted debugging. This loop repeats until the code passes all predefined checks.
- Secure Execution: All generated code is tested within an isolated, sandboxed environment to guarantee system safety and prevent unintended consequences during validation.