Large language models have advanced code generation and problem-solving, requiring comprehension of complex nacitural language tasks and generation of correct, efficient code. Current methods typically prompt LLMs to produce code in reasoning-step segments but remain prone to errors and struggle with complex real-world tasks. To address this, we propose VGHTCoder, a multi-agent code generation framework simulating the full programming lifecycle. It integrates Chain of Verification and Adaptive Debugging, combining programmer, code executor, and a verification agent with hypothesis testing to guide task completion. At each stage, the verification agent drafts responses, plans verification queries, and directs the executor, ensuring effective results. VGHTCoder outperforms single-agent and previous multi-agent models, achieving state-of-the-art pass@1 scores on HumanEval (94.5%), MBPP (92.17%), and MBPP-ET (64.2%), demonstrating strong potential for further advancement in code generation.