Anthropic engineer Addy Osmani: AI agents may make people “stop thinking independently”
On November 3, 2026, Addy Osmani, an engineer on the Anthropic Claude Code team, posted a long piece on X titled “Critical thinking when agents do the work,” saying that much of his daily work is now completed through AI agents, leaving almost no natural moments that force him to think critically. He explicitly said that he supports agentic workflows and is not calling for reduced use, but is emphasizing that “you have to deliberately schedule critical thinking into your day, because by default nothing will do that for you.”
Osmani described the relationship between people and agents as a spectrum with three stages: cognitive offloading, in which tasks are delegated but ownership of and judgment over the results are retained; deferral, in which the agent’s output directly becomes the next input and is no longer treated as a hypothesis to be verified; and cognitive surrender, in which the agent’s output is fully trusted, checking stops, and its answer quietly becomes one’s own conclusion. The term “cognitive surrender” comes from research by Wharton School professors Steven Shaw and Gideon Nave. He proposed a practical standard: if you cannot explain an answer, even with the help of an agent, then you have already given up ownership; low-risk tasks can be fully delegated, while high-risk tasks must retain judgment. Frequent surrender accumulates “cognitive debt.”
He suggested forming hypotheses before prompting, making verification a habit, and asking agents what they can and cannot verify. He also cited an internal Anthropic survey of 132 engineers, saying that “the skills needed to supervise Claude are precisely the skills that excessive AI use may shrink” (the supervision paradox). Business Insider and other English-language media reported this view around November 5.
Original sources and verification
Addy Osmani X long post (2026-11-03): https://x.com/addyosmani/article/2117512471930167377 (core quotations and structure verified)
Business Insider (2026-11-05): https://www.businessinsider.com/anthropic-engineer-addy-osmani-cognitive-surrender-warning-2026-11 (reporting framework verified)
Additional cross-checks: paraphrases by TechMoon, Shopifreaks, and others of the spectrum and habits are consistent with the original text; the source of the Wharton terminology is noted in the original.
ABAB AI Insight
Agentic workflows greatly improve execution speed and output fluency, while removing the friction that forces thinking from the workflow. In the past, critical thinking mostly happened in the “gap between the question and the answer”; once agents fill that gap, the default path becomes accepting fluent, seemingly correct output. The three-stage spectrum describes a gradual slide rather than a sudden break: each step seems reasonable (delegating, using the output directly, no longer checking), so it is hard to notice in oneself. Osmani’s position is practical rather than dismissive: ownership and explainability remain engineers’ bottom line; low-risk work can be offloaded, while high-risk work must retain hypotheses and verification. This is consistent with Anthropic’s own “supervision paradox”: the skills needed for oversight depend on skills that may deteriorate through excessive delegation, creating a potential mismatch between capability and responsibility. In the long run, cognitive debt reflects the externalization of system understanding, making it hard for individuals and teams to reconstruct explanations during later changes or failures.
ABAB News · Cognitive Laws
Once agents fill the gap between questions and answers, critical thinking shifts from a default behavior to a cost that must be deliberately scheduled.
The more fluent and correct the output, the weaker the motivation to check it; the boundary of ownership is whether it can be explained independently, not whether an agent was used.
The skills needed for supervision are precisely the skills that shrink first under excessive delegation — the mismatch between capability and responsibility accumulates in step with usage intensity.