Coding agents for large language models (LLMs) can today write software from natural language requirements, but the current code generation pipeline only verifies whether the generated code works and passes unit tests, not whether it satisfies the requirements it was supposed to implement. The consequence is a silent failure mode in agentic software development in which functionally-correct code silently violates some of the requirements, such as regulatory compliance, security constraints, or rate limits. In this paper, we present REAutonomy, a framework that pairs the development agent with a separate Requirements Engineering (RE) Testing Agent and a five-level, three-dimensional autonomy scoring framework (Decision Autonomy, Escalation Maturity, and Learning Autonomy) that evaluates how well the agent satisfies its declared requirements over multiple iterations. We provide a four-layer system architecture overview, a five-level test approach description, a JSONbased feedback loop mechanism, and an illustrative example in which a registration system requirement is iteratively improved from 0/15 initial autonomy score to 14/15 converged autonomy score over three iterations. We also present results from our implementation validation, including 106 automated unit and integration tests across six modules and ten-category end-to-end smoke test, which pass successfully. Finally, we contextualize our REAutonomy framework with recent requirement-driven code generation agents, and claim that an explicit, RE-driven compliance testing, coupled with a meaningful autonomy metric, is a missing component in the current agentic software engineering pipeline.

Artificial Intelligence