OpenAI’s board chair, Bret Taylor, says vibe coding—using AI to build software through natural language prompts—will remain a staple of development, but it’s not the final evolution of how we create and use software. In an episode of the "Big Technology Podcast" released on Wednesday, Taylor acknowledged that the ability to rapidly prototype apps with AI feels increasingly normal, but cautioned that focusing solely on speed misses the deeper transformation ahead. “Everyone’s looking at all the software use and saying, ‘How fast could I vibe code that?’” Taylor said. “I wonder if it’s the wrong question.” He argued that the real shift isn’t about building software faster, but about reimagining what software itself will become. Instead of traditional dashboards, web forms, and standalone applications, Taylor envisions AI agents as the future of software. These agents, he explained, will autonomously perform tasks by interacting directly with databases and systems. “We will delegate tasks to agents that will operate against a database,” he said. The critical question, he added, is who builds these agents—will they be purchased off the shelf, or developed in-house? While AI has dramatically reduced the cost of writing code, Taylor pointed out that it hasn’t solved the ongoing challenges of maintaining software, debugging errors, or ensuring long-term reliability. “That’s why most people would prefer to buy a solution off the shelf,” he said. “You want to amortize the cost of maintaining software among thousands of clients.” Despite its growing popularity, vibe coding has clear limitations. Google CEO Sundar Pichai praised the technology in a November interview, calling it “making coding so much more enjoyable” and enabling non-technical users to create simple websites and apps. He also noted that AI now generates over 30% of Google’s new code, up from 25% in October 2024. However, Pichai acknowledged that AI-generated code can be flawed—sometimes overly verbose, poorly structured, or riddled with security risks. “I'm not working on large codebases where you really have to get it right, the security has to be there,” he said. Similarly, Boris Cherny, the engineer behind Anthropic’s Claude Code, has warned that vibe coding works best for early prototypes or temporary projects, not for mission-critical systems. “You want maintainable code sometimes. You want to be very thoughtful about every line sometimes,” he said in a December episode of “The Peterman Podcast.”