New Framework Lets Non-Programmers Turn Requirement Documents Into Working Software

Researchers have introduced ARC (Agentic Requirement Compilation), a framework designed to bridge the gap between software requirements and executable code by automating the translation of complex, multi-modal requirement documents into fully functional web systems.

The Challenge It Addresses#

Large language models have made coding more accessible by converting natural language into code, but they struggle when faced with large-scale specifications containing hundreds of scenarios. These models frequently produce flawed implementations or overlook essential constraints. ARC tackles this problem by treating requirement documents as computable specifications that can be systematically transformed into working software.

How ARC Works#

Rather than relying on a single translation pass, ARC employs a bidirectional workflow combining two complementary approaches:

• Top-down architecture phase: The system first designs the modular structure of the application
• Bottom-up implementation phase: Code is then generated to match automatically created test suites

This test-driven strategy ensures that generated implementations satisfy their specifications. The framework also produces comprehensive test coverage and maintains traceability linking requirements to design decisions to final code.

Performance and Accessibility#

Testing across six web system benchmarks and 101 mobile app tasks from the AppForge benchmark revealed substantial improvements. ARC achieved a 100% compilation success rate, passed 68.3% of tests on AppForge, and performed 50.6% better on average for GUI tests compared to existing language model approaches.

In a user study with 21 participants, people with minimal programming experience successfully wrote requirement documents describing up to 174 scenarios—completing the process in roughly 5.6 hours on average. One generated system, a real-world ticket-booking application, reached approximately 10,000 lines of production-quality code. This outcome suggests the framework could democratize software development by enabling non-technical stakeholders to specify complex systems that automatically compile into maintainable, runnable applications.