Recent advances in Large Language Models (LLMs) have enabled the emergence of coding agents capable of software development, debugging, refactoring, and operational maintenance. Most modern agents employ variants of the ReAct (Reasoning and Acting) paradigm, in which an agent alternates between reasoning and action while observing environmental feedback. However, practical engineering experience reveals a fundamental limitation of ReAct-based coding agents: reasoning is probabilistic, while software systems require deterministic correctness. As a result, coding agents frequently perform actions based on unverified hypotheses, leading to unnecessary code modifications, hallucinated root-cause analyses, over-refactoring, and cascading system failures. This paper proposes Evidence-Gated ReAct (EGR), an engineering framework that introduces a Verification Layer and an Evidence Layer between reasoning and action. The central principle of EGR is simple: Reasoning is not evidence. Actions require evidence. By enforcing evidence collection prior to code modification, EGR transforms assumption-driven development into evidence-driven engineering. We further define four evidence models applicable to software engineering: Trace Evidence, Acceptance Evidence, Regression Evidence, and Performance Evidence. The framework provides a practical methodology for improving reliability, explainability, and operational safety in AI-assisted software development.