这篇2026年2月26日的报道追踪了一次反转：由前OpenAI研究员Andrej Karpathy在2025年2月提出的“vibe coding”——即基于聊天的编码——已与压力绑定。Anthropic的Claude Code和OpenAI的Codex等AI编码代理现在可以规划、执行并完成工作，因此高管越来越把持续使用视为强制要求。除了文本或图像生成外，这些系统在很少人工监督下运行任务，Greg Brockman声称代理的空闲时间是“浪费机会”的说法，推动了持续生产力压力的常态化。
加州大学伯克利分校的一项200人组织研究显示，即便工作正在被AI代理接管，员工仍在工作更长时间。工程师们报告“AI fatigue”和害怕错过突破性提示机会。DocuSketch的Andrew Wirick称，20分钟试用Anthropic的Opus 4.5便生成了可运行原型，并抬高了预期；该团队追踪工程师的“每日报告互动次数”与每周低效循环；而Arcade.dev的CEO Alex Salazar监控Claude Code账单并推动更高使用率。在一次“come to Jesus”会议后，他的团队账单 reportedly上升了10倍。
高管自己的反复试验进一步放大压力：他们自己构建演示并为所有人树立预期。然而Section的报告显示了差距——超过40%的C-suite领导者称AI每周可节省至少8小时，而67%的非管理者称节省低于2小时或无节省。Intuit让产品经理和设计师在QuickBooks中vibe code功能，Balazs称工程交付速度最高提高30%。但伯克利研究者将其称为“task expansion”：工程师还要清理非技术同事的产出，并承担由角色模糊带来的额外工作量。可能结果是“busyware”——低价值、一次性代码——而关键的生产力策略变成了知道什么不应建。
The Feb. 26, 2026 report tracks a reversal: “vibe coding,” introduced by former OpenAI researcher Andrej Karpathy in February 2025 as chat-based coding, has become tied to pressure. AI coding agents such as Anthropic’s Claude Code and OpenAI’s Codex now plan, execute, and complete work, so executives increasingly treat continuous use as mandatory. Beyond text or image generation, these systems run tasks with little human oversight; Greg Brockman’s claim that idle agent time is “a wasted opportunity” helps normalize nonstop productivity pressure.
A University of California, Berkeley study of a 200-person organization shows that even while work is being offloaded to AI agents, employees are still working longer hours. Engineers report “AI fatigue” and fear of missing breakthrough prompt opportunities. At DocuSketch, Andrew Wirick says a 20-minute Opus 4.5 test produced a working prototype and raised expectations; the team tracks engineers’ “interactions per day” and weekly inefficient loops, while Arcade.dev CEO Alex Salazar monitors Claude Code bills and pushes higher usage. After one “come to Jesus” meeting, his team’s billing reportedly rose 10-fold.
Executives’ own tinkering further amplifies pressure: they build demos themselves and set expectations for everyone. But Section reports show a gap—over 40% of C-suite leaders say AI saves at least 8 hours weekly, while 67% of non-managers say savings are under 2 hours or none. Intuit lets product managers and designers vibe code QuickBooks features, and Balazs cites up to 30% higher engineering velocity. Yet Berkeley researchers call this “task expansion”: engineers also clean up nontechnical colleagues’ outputs and absorb extra workload from blurred roles. The likely result is “busyware”—low-value, disposable code—and the key productivity tactic becomes knowing what not to build.