The Agent Can Type All Day. The Job Was Never Typing.
July 6, 2026 · Marcus Webb, Dark Time~5 min read
You've felt it. The agent's cranking out code, the diff is stacking up, and you'd swear you're moving twice as fast as you were last year. Then the sprint ends and somehow the same amount of stuff got done. That gap — between how fast it feels and how fast it actually went — isn't in your head. Somebody finally went and measured it, and the number's a little embarrassing.
The survey nobody wanted to publish
A team surveyed 2,989 developers — pulled from about 8,000 engineers at BNY Mellon — about writing code with GitHub Copilot. Eighty-six percent said they were satisfied. Happy with the tool. Then the same survey asked how much time it actually saved them, and roughly 60% said less than an hour a week. Less than an hour.
Here's the part that stuck with me. When they lined up "how satisfied are you" against "how much time did it save," the two barely moved together — a correlation of 0.34, which in plain terms means the thing that makes you love the tool is mostly not the thing that saves you time. The paper's called Beyond the Commit, out this year, and they backed the numbers with interviews of eleven of those developers. Zoom out and it gets worse: an NBER paper this February surveyed around 6,000 executives, and more than 80% of their firms reported AI had made no measurable dent in productivity over three years. No measurable dent.
So where does the time actually go?
Here's the thing about writing code: you barely do it. One of those engineers put a number on it — when you're refactoring, 70 to 80% of the time is reading the code, and maybe 10 to 20% is actually typing new lines. The typing was never the slow part. It's the same as any repair job. Ever watch someone chase a weird electrical fault in an old house? They spend three hours with a meter and a flashlight tracing which wire does what, and ninety seconds with the screwdriver. The screwdriver was never the bottleneck. The knowing-which-wire was.
And that's exactly where the agent comes up short. Same survey: the folks who handed the agent a whole task and said "go" did worse than the folks who chopped it into small pieces and fed them one at a time. One engineer said the AI "typically fails at code improvement without a lot of context and very specific instructions." Another said the reason the experienced people still had the edge, when something broke in production, was that they "know exactly where the issue is." The tool can write the fix. It just doesn't know where to point it.
An old book already named this
Twenty-odd years back a Chinese programmer named Liu Weipeng wrote a book, Dark Time, mostly about how people actually learn and think. One idea runs through the whole thing: the knowledge that makes you good at something doesn't sit in a file you can email. It's a network in your head — a web of connections you built by reading, getting burned, and reading some more. He's got a line I keep coming back to: if you can't teach it clearly, you probably don't really understand it. Flip that around and you've got the whole problem with the agent. Teaching it — writing the spec clear enough that a thing with no memory of your codebase can act on it — is the same act as understanding it yourself. The stuff you can't write down clearly enough to hand off has a name: tacit knowledge. And it's not some small leftover. In real programming it's most of the job.
Refactoring is 70–80% reading code and only 10–20% writing (developer interviews in "Beyond the Commit," 2026); the agent takes the writing, but the tacit project knowledge — knowing where the issue is — stays in your head. Framework: Liu Weipeng, Dark Time (knowledge as a network you can't hand off). One engineer's read of a fast-moving field.
What this means for you on Monday
Stop handing the agent the whole task and hoping. The survey already ran that experiment for you, and the people who did it came out behind. Chop the work into pieces small enough that each one has a clear, written spec — because that's the only part the agent can actually take. And notice what you're doing when you write that spec: you're dragging the thing out of your head and making it explicit, which is the exact work Liu calls the real thinking. His other line is that writing something down is how you think it through — not a record of thinking you already finished. Same here. The spec you write for the agent is you figuring out the problem. The agent just types up the answer you already found.
You can only hand off the part you can write down
Most of real programming is reading and knowing where the issue lives — a network in your head you can't email. The agent gets the typing; the knowing stays with you.
So no, the tool isn't a fraud, and the 86% aren't wrong to like it. It's genuinely nice to not type the boilerplate. But "nice" and "faster" turned out to be two different measurements, and the survey caught them coming apart. The work that's left after the agent takes the typing is the work that was always the actual job — reading the code, holding the map in your head, knowing which wire to pull. That part didn't get automated. It got exposed.
Framing drawn from Liu Weipeng, Dark Time (knowledge as a network in your head; if you can't teach it clearly you don't really understand it; writing as thinking). Survey and interview figures are from "Beyond the Commit: Developer Perspectives on Productivity with AI Coding Assistants" (2026) and an NBER executive survey (February 2026). This is a fast-moving field and one working engineer's read, not a verdict on any specific tool — plenty of teams do get real gains.
AI 差的正是这一口。还是那份问卷:把一整个任务丢给 AI、说句「你去干」的人,反而干得不如把活儿拆成小块、一块一块喂进去的人。有位工程师说,AI「在没有大量上下文和非常具体的指令时,通常搞不定代码改进」。另一位说,出了线上问题,老手还占优,就因为他们「清楚知道问题在哪儿」。工具能把补丁写出来,可它不知道该往哪儿打。
有本老书早就给这毛病起过名
二十多年前,有个中国程序员刘未鹏写过一本《暗时间》,大半在讲人到底是怎么学、怎么想的。有条线贯穿全书:让你真正厉害的那点本事,不装在一个能发邮件发出去的文件里。它是你脑子里的一张网——靠一遍遍读、一次次踩坑、再接着读,一点点连起来的。他有句话我老想起来:你要是讲不清楚,那多半是你自己还没真懂。反过来一想,这不就是 AI 的死穴嘛。教会它——把需求写清楚到一个对你代码库毫无记忆的东西也能照着干——这件事本身,就跟「你真搞懂了」是同一件事。那些你没法写清楚、交不出去的东西,有个名字叫隐性知识;它可不是剩下的一点边角料,在真实的编程里,它就是活儿的大头。
重构里 70%–80% 的时间在读代码,真正写代码只占 10%–20%(《Beyond the Commit》2026 的开发者访谈);AI 接走了「写」,可隐性的项目知识——知道问题在哪儿——留在你脑子里。框架:刘未鹏《暗时间》(知识是一张交不出去的网)。一位一线工程师对快速演进领域的解读。
这事儿周一上班对你意味着啥
别再把一整个任务囫囵丢给 AI,然后干等着了。这个实验问卷替你做过了,这么干的人是落在后头的。把活儿拆到每一块都能配一份写清楚的需求——因为那一块,才是 AI 真正接得住的。还有,你留意一下你写那份需求时在干嘛:你是在把脑子里那张网往外掏、把它摊成明文,而这一步,恰恰就是刘未鹏说的、真正的思考。他还有句话:写下来不是把想好的记个账,写本身就是把事想清楚的过程。这儿一模一样。你给 AI 写的那份需求,就是你在把问题想明白;AI 不过是把你已经找到的答案敲成字。
框架取自刘未鹏《暗时间》(知识是你脑子里的一张网;讲不清就是没真懂;写作即思考)。问卷与访谈数字出自《Beyond the Commit:开发者眼中的 AI 编程助手与生产率》(2026)及 NBER 一份高管调研(2026 年 2 月)。这是个快速演进的领域,也只是一个一线工程师的看法,不是给哪个具体工具下判决——不少团队确实拿到了真实的收益。