Lazy, Considerate, and Unshakable: The Three Human Traits AI Agents Make Valuable

Three Kinds of People Who Thrive in the AI Era: Lazy Thinkers, Considerate Communicators, and Those Who Aren't Led Astray by AI

Kent Beck said in 2023 and repeated in a 2025 interview: "90% of my skills are now worthless, and the remaining 10% has increased in value by 1000 times." That 10% is figuring out what to do, setting milestones, and controlling design complexity.

Two years later, that sentence has become everyday reality in many companies. Boris Cherny, head of Claude Code, said in January that all of his code has been written by Claude for over two months; he merged 22 PRs the previous day and 27 the day before that. When execution becomes cheap, which human traits become valuable? My judgment is three kinds of people: those with ideas but who are lazy to act, those who habitually think about others, and those who have judgment and won't be misled by AI. The first two were often seen as character flaws; now they are exactly advantages. The third determines whether the first two go in the right direction.

Lazy People with Ideas: Agents Fill in the Execution Gap

Larry Wall, author of Perl, listed laziness as the first virtue of a programmer in 1996. His definition: willing to invest significant effort upfront to reduce overall energy expenditure. Lazy people aren't doing nothing; they just don't want to do the same thing twice.

In the past, people with ideas but no willingness to execute were at a disadvantage. Ideas stayed as talk; either they asked others for help or struggled along slowly by themselves. Most ideas died that way. Agents fill exactly that gap. They can work for hours straight, look at logs when tests fail, fix code, and run again without getting tired or annoyed. What they lack is direction: they won't proactively think "this product isn't good" or suggest "try a different approach."

Cherny described his current work: he can think about what to do next, do creative things, and leave details to the model. An Anthropic article from July 2025 put it more bluntly: Claude Code lets "anyone who can describe a problem clearly build the solution." Their examples included a legal team building their own system.

You can be lazy about execution, but not about ideas. Ideas are the only thing these people must produce personally.

Considerate People: Think for Users, Think for Agents

Being considerate has two layers: one for users, one for agents.

Start with users. Andrew Ng wrote in July 2025 that as coding cost drops, the bottleneck shifts to product management: deciding what to build. His view: product managers with strong empathy for users can often make correct decisions intuitively because they have an accurate user model in their heads. AI loyally executes your perspective: if you stand at the boss's angle, it builds a product full of conversion entry points; if you stand at the user's angle, it builds a product people want to stay on. That judgment, the model cannot make for you.

Then for agents. Anthropic's official prompt documentation advises treating Claude as "a very smart new employee who doesn't know your rules and processes." The documentation's test is simple: show your prompt to a colleague who doesn't know the task; if they can't understand it, neither can Claude. Anthropic's engineering team also wrote about "getting into the model's shoes" when building agents. When building code agents, they spent more time polishing tools than prompts; agents kept using wrong relative paths after leaving the root directory, so they changed the tool to force absolute paths, and the problem disappeared. In another multi-agent research system, they had one agent rewrite a flawed tool description, and subsequent agents using that tool finished tasks 40% faster.

Thinking for agents, in practice, means saying things clearly in one go. In June 2025, Shopify CEO Tobi Lütke proposed replacing "prompt engineering" with "context engineering." His definition: give all the context needed for the task so that it becomes solvable for the model. Karpathy publicly agreed.

This is exactly the habit of "people who don't like to bother others." Before asking for help, they think through the goal, constraints, and what counts as done, and communicate it clearly the first time, so the other side doesn't have to ask back and forth. OpenAI's GDPval paper had a comparison: trimming task descriptions to 42% of the original length made model performance noticeably worse. Ethan Mollick's January article put it more directly: what's scarce now is knowing what to ask for; people who already know how to write requirements documents and assign tasks to subordinates write prompts that work well as-is.

Note that giving enough context isn't the same as being verbose. Anthropic, when discussing context engineering, emphasizes finding "the minimal set of high-signal information"; the more you stuff in, the easier it is for the model to miss the point. Considerate people may not talk much, but every sentence counts.

People with Judgment: Not Being Led Astray by AI

The first two kinds are about how to delegate work. The third is about how to evaluate what comes back.

The easiest way AI misleads people is that its output always looks correct: code runs, documents are fluent, tone is confident. METR did a randomized controlled experiment in 2025: 16 senior open-source developers working on their own codebases took 19% longer when allowed to use AI; they expected to be 24% faster and still felt 20% faster afterward. The experiment used early-2025 models and tools, so conclusions can't be directly extrapolated, but it shows one thing: even senior engineers misjudge how much AI helps, let alone the quality of its output.

People with judgment put their own judgment first. Anthropic published an internal survey in December 2025; many engineers said they had become "managers of AI agents," and some said over 70% of their work is now reviewing and fixing code. But in the same survey, more than half said work that can be fully delegated to Claude accounts for only 0 to 20% of tasks; what they delegate is mostly tasks that can be checked for correctness at a glance, while high-level thinking and design stay with them. They don't distrust AI; they know which conclusions they can verify and which they can't.

Simon Willison calls this ability "vibe engineering." His list includes automated tests, planning ahead, a code review culture, and solid manual testing, plus a specific intuition: knowing what can be outsourced to AI and what must be handled yourself.

One more true line from the same Anthropic survey: the more you want to do a task yourself, the less likely you are to hand it to Claude. People keep the work they enjoy. That's fine. What matters is that for what you delegate, you can judge right from wrong, and the direction always stays with you.

What These Three Have in Common

Trait	Previous situation	Working with agents
Lazy with ideas	Ideas never land	Ideas stay with you, execution goes to agent
Considerate	Think too much, act slowly	User perspective sets direction, give context in one go, agent executes fast and accurately
Has judgment	Always important	Judging AI output, holding direction becomes the main job

Back to Kent Beck's sentence. The devalued 90% is execution-level proficiency; the rising 10% is knowing what to do, speaking clearly for users and agents, and judging whether the returned results are right. Those three things, agents can't replace.