Addy Osmani works on the Claude Code team at Anthropic, and he says his own work has reached a point where nothing forces him to think. "Increasingly, nothing in my day-to-day forces critical thinking on me now that so much of my work runs through agents," he wrote in a Tuesday X essay, as Business Insider reported on November 5, 2026. His answer is not to drop agents. It is to build thinking back into the day on purpose.
We could not open the X essay itself. This post uses the Business Insider summary of it, plus Osmani's longer earlier essay on cognitive surrender on his own site, dated May 5, 2026, which lays out the same argument in full.
TL;DR
| Question | Answer |
|---|---|
| Who is warning? | Addy Osmani, on the Claude Code team at Anthropic |
| The core claim | When agents do the work, critical thinking no longer happens by default; you must build it in |
| Offloading vs surrender | Offloading keeps ownership of the answer; surrender is adopting the AI output with nothing left to check |
| Key evidence cited | Wharton study: 73% accepted wrong AI answers across 1,372 participants |
| His position on AI | Keep using it; know which mode you are in |
| Test for ownership | "When you can't explain the answer to a question, even by way of your agent, you've ceded ownership" |
What are the three kinds of diminished thinking?
Business Insider reports that Osmani describes three patterns.
- Cognitive offloading. You delegate tasks to AI while retaining ownership of the product. "I do this all day," he wrote. "You should, too."
- Deferral. The AI's output becomes your new input.
- Cognitive surrender. Complete trust in AI output. The term was coined by two Wharton researchers, and Osmani calls it lazy. Surrendering often builds what he calls "cognitive debt."
He is explicit that stakes matter. "Where the stakes are low, feel free to delegate entirely." And on the line between help and offloading: "Both are OK, as long as you know which one you're doing."
Where does the Wharton evidence come from?
In his May essay, Osmani credits the term to a paper by Steven Shaw and Gideon Nave at the Wharton School: "Thinking - Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender."
As Osmani summarizes it, across three experiments with 1,372 participants, simply having an AI available was enough for people to surrender. On trials where the AI was wrong, participants accepted the wrong answer 73% of the time. Their confidence also rose with AI available, even though half the answers were deliberately incorrect. He describes this as "borrowing the model's confidence."
We have not read the paper ourselves, so treat the figures as Osmani's summary of it. Shaw is quoted in the essay as warning against alarmism: the issue is calibration, "knowing when AI is helping you think and when it is quietly doing the thinking for you."
For other studies on the same theme, see our coverage of the PsyArXiv study on AI advice and "I don't know", the Berkeley and CMU persistence study, and our explainer on whether AI makes you dumber.
Where does surrender show up in software work?
Two shapes reviewing one sheet, one marking a green check, showing code review of agent output and cognitive surrender
Osmani lists places he has seen it, "mostly to me":
- Reading the diff. An agent produces a 600-line pull request. Names look reasonable and tests are green. You approve. A subtle change to a transaction boundary slips by. "You didn't review the code. You ratified it."
- Debugging an error you don't understand. You paste the stack trace, the fix works, you move on. Two weeks later a related symptom returns, and you never understood the original bug.
- Making a design call. The agent picks a queue over a direct call with a confident paragraph. You never reasoned about throughput, failure modes or replay.
- Learning something new. He cites the Anthropic skill-formation paper: engineers who used AI to generate code while learning a new library scored 17% lower on a follow-up comprehension quiz than a control group. Engineers who used AI for conceptual inquiry held their ground. "Same tool. The posture changed the outcome."
He argues engineers are unusually exposed because generated code compiles, lints and looks like the rest of the file; because throughput is the visible metric; because models speak in declaratives that reviews read as authority; and because surrender compounds, since each accepted chunk makes the next change another act of surrender.
How does this connect to comprehension debt?
Osmani ties surrender to comprehension debt, the gap between how much code exists and how much of it a human understands. His line: cognitive surrender is "the mechanism by which comprehension debt accumulates." Each act is a tiny loan. He pairs it with the MIT "Your Brain on ChatGPT" work on cognitive debt, which he says found reduced neural connectivity and weaker recall among writers leaning on AI. Again, those are his summaries.
Our post on the debate over retyping LLM code covers a hands-on take on the same debt.
What does he suggest doing about it?
From the May essay, the personal habits:
- Form an expectation first. Before running the agent on a non-trivial task, write down what you think the answer looks like. If the agent differs, you have a real choice: are you wrong or is it?
- Review the diff like a junior wrote it. "Seems right" is still not a review.
- Ask the model to argue against itself. A cheap second pass breaks the borrowed-confidence effect.
- Notice when you are tired. The first PR of the day gets a real review; the fifth gets a glance.
- Check where your confidence comes from. If you cannot rebuild why a design was chosen, rebuild it before the meeting continues.
The structural habits:
- Verification as an exit criterion. Every agent task ends in evidence: a test that runs, a screenshot, a log or a reviewer sign-off. Business Insider adds he told engineers to ask the agent what it could verify and what it could not.
- Anti-rationalization tables. Pair each excuse for skipping a step with a written rebuttal.
- Small PRs. "The unit of review is the unit of comprehension."
- Ask before you generate when learning. Explanations build the mental model; generation can erode it.
And from the Business Insider piece, the questions he asks himself, including "Did I formulate a hypothesis before asking?" and "What am I no longer practicing?" His warning on skills: "Skills that you stop exercising are going to go away, and probably faster than you might guess." His advice: "Keep doing the ones you're not willing to let go."
A checklist you can paste into your agent workflow
A green thread rising from an open book toward stacked blocks, showing learning through conceptual inquiry with AI agents
This is our adaptation of his points, not his wording.
| Step | Prompt or action |
|---|---|
| Before | Write three lines on the answer you expect |
| During | Ask: "What can you verify, and what can't you?" |
| After | Ask: "Argue the opposite design and tell me which assumptions decide it" |
| Review | Cap diffs near 100 lines or split them |
| Exit | Attach evidence (test, trace, screenshot) to the task |
| Weekly | Name one skill you are no longer practicing and schedule it |
If you run coding agents at work, our guides to Claude Code habits and the token efficiency guide cover the tooling side. For the person behind the essay, see Osmani joining Anthropic's Claude Code team.
What is open or contested?
- The new X essay is not independently read by us. Our account of it comes from Business Insider. The details of the May essay are first-hand from his site.
- The evidence is early. Business Insider notes it is not yet clear whether AI is making people less capable or changing how they think, and the researchers it spoke with cautioned against jumping to conclusions.
- The Wharton numbers are as summarized by Osmani. Check the paper before citing them.
- An Anthropic employee has an interest. Osmani helps build the product the essay is about. He says plainly that he is not against it, and that the habit matters more than the tool.
Bottom line
The practical point is small and hard to do: decide, each time, whether you are offloading or surrendering. If you cannot explain the answer, even through your agent, you do not own it. Write your expectation first, ask for evidence, and keep one skill in practice.
Quotes are as reported by Business Insider on November 5, 2026 and on Osmani's site on May 5, 2026.