SpecFirst: Behavioral Specification Elicitation as a First-Class Step in Agent-Based Program Synthesis from Scratch

SpecFirst decomposes from-scratch program synthesis into two explicit stages. A specification agent first probes an execute-only binary and combines observations with documentation into a structured behavioral specification. A code synthesis agent then implements against that stable reference. Across all 200 ProgramBench instances and four models from two capability-diverse families, the framework improved test pass rates by 6.9%–21.3% and binary exploration coverage by 9.4%–18.5% over a single-loop baseline, with statistically significant results. The paper argues that requirements engineering should be an explicit component of agentic coding systems.