Imagine a team of AI agents that not only write software but also learn from their mistakes and improve their collaboration in real-time. This isn't science fiction; it's the reality of EvoMAC, a groundbreaking new approach to automated software development. Traditional AI struggles with the complex, multi-step process of building software. Single AI models often lack the reasoning skills and long-term planning needed for large projects. EvoMAC tackles this by creating a network of specialized AI agents that work together, much like a human development team. One group of agents, the 'coding team,' focuses on writing the code itself, while another, the 'testing team,' designs unit tests to rigorously check for errors. The real magic happens with the 'updating team.' Inspired by how the human brain learns, EvoMAC incorporates a feedback loop. When the testing team finds bugs, this information isn’t just presented as a report; it's used to automatically update the coding team’s instructions. This ‘textual backpropagation’ allows the agents to learn from their errors, refining the code and their collaboration with each iteration. To push the boundaries of software-level AI coding, researchers developed rSDE-Bench, a new benchmark designed to test AI's ability to handle real-world software development challenges. This benchmark includes detailed requirements and automatic evaluation, making it possible to accurately assess the AI's performance. Experiments show that EvoMAC significantly outperforms other state-of-the-art methods on both function-level and software-level tasks, showcasing its effectiveness and potential for transforming how we build software. The future of software development could be far more automated and efficient, thanks to evolving AI teams working together seamlessly.