AI News Daily한국어

Can you explain the code your agent wrote? Claude Code's Addy Osmani on the line between offloading and 'surrender'

Nov 15, 2026Original Addy Osmani(앤트로픽) X 에세이 · Business Insider · 와튼스쿨 · 마일드코드 ↗

In 3 lines

  • Osmani splits AI use into offloading, deferral and cognitive surrender
  • In a Wharton study, wrong AI answers cut accuracy from about 46% to 31%
  • Delegate where being wrong is cheap; an answer you can't explain isn't yours
Can you explain the code your agent wrote? Claude Code's Addy Osmani on the line between offloading and 'surrender'

Addy Osmani, who builds Claude Code at Anthropic, posted a long X article titled "Critical thinking when agents do the work" on Nov. 3. It opens with an admission that nothing in his day forces critical thinking on him anymore now that most of his work runs through agents, and sorts how we hand work to AI into three stages: offloading, deferral and surrender. In twelve days it drew 5.6 million views, 21,000 likes and 51,000 bookmarks. We read it alongside a Wharton experiment, Anthropic's own research and a Korean YouTuber who put the idea to work.

"Nothing in my day-to-day forces critical thinking on me"

Osmani spent years on Google's Chrome team before moving to Anthropic this year to work on Claude Code. He opens by admitting that "nothing in my day-to-day forces critical thinking on me now that so much of my work runs through agents." He addresses engineers and product people running many agent sessions in parallel, and says plainly that he is a big fan of these tools and is not arguing for less agentic work. The space between question and answer, where most of our thinking used to happen, "is now measured in seconds."

Offloading, deferral and surrender

He describes a spectrum. At the top is cognitive offloading, delegating to the AI while keeping ownership of the result: "I do this all day. You should, too." In the middle is deferral, where the tool's output becomes your next input and stops being a hypothesis. At the bottom is cognitive surrender, where the AI's output quietly becomes your own and you feel no need to check it. You read the last answer well, so you skim the next one and scan only the skeleton of the one after, until someone asks about a line and you can't explain it, even with your agent's help.

When the AI was wrong, accuracy fell from 46% to 31%

"Cognitive surrender" was coined by Steven Shaw and Gideon Nave at Wharton. They gave 1,372 people 9,593 reasoning puzzles with an AI chatbot on hand. When the AI gave correct advice, accuracy rose to 71% from about 46% unaided, but when it gave faulty advice accuracy fell to roughly 31%. In the first study about 80% of those who opened the chat followed the wrong advice, and even a bonus plus instant feedback only raised the rate of rejecting it from 20% to 42%. "The key issue is calibration," Shaw said: knowing when AI is helping you think and when it is quietly doing the thinking for you.

Steven Shaw and Gideon Nave · Wharton · PsyPost ↗

Better models make it harder

Osmani argues the problem gets harder as models improve, because obvious errors become rarer and most outputs are well written and mostly correct, so they don't demand a second look. An agent hands you a 600-line PR; you scan the structure, see reasonable names and green tests, and approve, while somewhere in the middle a transaction boundary's ordering has shifted or an edge-case default has flipped. "Even if you are running many agents at a time, you still only have one pair of eyes on the outputs," he writes. "When you're producing more than you can stand behind, you're no longer reviewing. You're hoping."

You don't need to know everything; the stakes decide

He is not asking anyone to hold a whole modern codebase in their head. The Claude Code team, he notes, rebuilds understanding on demand by asking Claude to explain the slice it needs. He let an AI pick one of three similarly rated hotels for a course recording and would do it again. He would not do the same for how to invest his money, how to advise his kids on careers, or at work for what to build and whether it's safe to ship: "Take the handoff where being wrong is cheap, keep the reasoning where others will have to deal with the consequences."

Addy Osmani · Claude Code team, Anthropic · Business Insider ↗

Form a hypothesis before you ask

These are the questions he now asks himself: Did I formulate a hypothesis before asking? Am I asking it to help me think, or to think for me? What would have to be true for this to be wrong? He also asks whether he has heard the counterargument from another model, whether he would bet on the answer in front of the person who has to maintain it, and what he is no longer practicing. He writes the guess down first, even a bad one, because once you've read a well-written response it is hard to imagine any other possibility.

What did the green check prove?

Verification can be surrendered too. "Is there evidence that supports my conclusion?" quietly becomes "Did the check turn green?", and a test written with the same mistake as the feature will pass. So he defines "done" up front as something tangible, such as a passing test, a screenshot or a log, and reads the command the agent ran and its output before the summary. Asking "Did I miss anything?" after a complex feature usually surfaces two or three more things, he says, and he likes a habit of having Claude write a change report with a quiz at the end: merge only after a perfect score.

An agent won't stop you

A vague request used to get challenged by a colleague, and that was when you realized you hadn't figured it out. An agent fills in what you left out and hands back something that looks done, so Osmani says to write your own brief first: the audience, what is out of scope and the definition of done. Korean Claude Code instructor Mildcode tried this in a Nov. 12 video. With the project's mission and a "won't do" list written to a file, a one-line request for a savings-recommendation feature made Claude Code stop and offer three options. His advice: say "give me options" instead of "just handle it."

Mildcode · YouTube ↗

"Supervising Claude requires the very coding skills that may atrophy"

To the objection that models will soon catch what he can, Osmani answers that "regardless of how much AI helped, a human has to take responsibility." Anthropic's December 2025 study of its own engineers calls this the paradox of supervision: supervising Claude requires the very coding skills that may atrophy from AI overuse. In an experiment Anthropic published in January with 52 developers learning a new library, the AI group averaged 50% on a comprehension quiz versus 67% for those coding by hand, while participants who used AI only for conceptual questions scored 65% or higher.

Anthropic · Anthropic Research ↗

Once a week, without an agent

The habits he suggests are small. Once a week, write a bit of code without an agent or do one real thing that would hurt to forget; once a month, write down what you've stopped doing and decide which of those you're fine letting go. If something is above your comprehension, go learn it rather than pasting in a one-line explanation. "Skills that you stop exercising are going to go away, and probably faster than you might guess," he concludes. "Keep doing the ones you're not willing to let go."

OriginalAddy Osmani(앤트로픽) X 에세이 · Business Insider · 와튼스쿨 · 마일드코드↗

More insights