Type: url
Original: https://arxiv.org/pdf/2407.15568
Status: generated
Summary
The paper proposing AgileGen, an Agile-based approach to generative software development through human-AI teamwork, from Tianjin University and CSIRO’s Data61. AgileGen uses the Gherkin language of behavior-driven development to turn an end user’s brief requirement into user scenarios with acceptance criteria, converts them to natural language for the user to confirm, add, delete or modify, and then uses the decided scenarios to drive visual design, “consistency factors” and code generation, followed by iterative acceptance and recommendation by the user; a memory pool stores past users’ scenario decisions for reuse. Evaluated on web projects from “50projects50days” and a sample of ChatDev’s SRDD dataset, it reports better CodeBLEU and Pass@1 than the compared agents and high user satisfaction.