Since May, the author has experienced a significant evolution in “vibe coding,” enabling rapid code deployment. The shift from manually crafting code to relying on AI agents has transformed their workflow. Initially, they found it surprising that prompts could generate functional code, but this has become an expectation. They emphasize that while coding requires some hard thinking, most software development mainly involves transferring data between formats. The author mainly uses Command Line Interfaces (CLI) as a starting point for apps, which agents can verify for output. The release of GPT-5 marked a turning point. The author describes how their trust in the model has grown, allowing them to focus less on reading code and more on high-level structures and dependencies. They prefer TypeScript for web applications, Go for CLIs, and Swift for macOS development. The comparison between Codex and Opus reveals that Codex excels in understanding and reading existing code, which can lead to more efficient outcomes, despite sometimes taking longer to respond. The author appreciates Codex’s conversational approach to coding, contrasting it with the older planning methods that felt cumbersome. The introduction of Oracle, a CLI tool that enhances GPT-5's capabilities, has streamlined their workflow, allowing for quicker interactions. The author’s own projects, like VibeTunnel, illustrate how AI can now handle complex coding tasks that previously required significant manual effort. Now, with GPT-5.2 released, the author notes a marked improvement in performance and a more up-to-date knowledge base compared to its predecessors. They’re currently focused on a project called Clawdis, an AI assistant with extensive control over their digital environment, showcasing the continued integration of AI into their development process.