The author reflects on the rapid evolution of programming tools and the increasing capabilities of AI agents, particularly in coding. A year ago, they were excited about the potential of LLMs after the release of Claude Code, but now they find that the models have improved significantly. For instance, a model like Opus can write 90% of their code, a stark contrast to the 25% that Claude Code could manage just a year prior. Despite public benchmarks being unreliable, the author sees qualitative advancements in coding models, leading to significant shifts in their own programming habits—from spending 80% of their time reading code at a large company to just 5% in writing code at a startup. The author reminisces about their experiences with IDEs, noting that while they once felt certain about the superiority of integrated development environments, they now prefer to use simpler tools like Vi. They emphasize that working with agents involves understanding their limitations and constantly adapting. The author is building a new tool, exe.dev, to create an environment where agents can operate without the constraints that often come with existing solutions. They express a newfound joy in programming, as agents allow them to explore and create more efficiently. Despite the optimism about AI in programming, the author acknowledges the fears surrounding job displacement and technological change. They draw parallels to historical shifts in labor, like the decline of agricultural jobs, suggesting that while change brings challenges, it can also lead to progress. They criticize the shortcomings of current software products, such as Stripe Sigma, stating that they often don't meet their needs. Instead, they rely on agents to perform tasks more effectively. The author concludes with a philosophy that prioritizes the needs of programmers over traditional customer-driven approaches, advocating for better tools that align with how programmers actually work.