Article · AI translation
IT Home reported on November 4 that Addy Osmani, an engineer on Anthropic’s Claude Code team and a former Google technology executive, published a lengthy article on X titled “How to Stay Critical When AI Agents Do the Work.”
He believes that as AI agents become increasingly powerful, people may gradually lose the ability to think independently, verify results and understand complex systems while growing more reliant on AI, ultimately treating AI’s judgments as their own. He calls this phenomenon “cognitive surrender.”
Osmani worked at Google for more than 14 years, overseeing Chrome developer experience and Google Cloud AI related work. He now works on Claude Code at Anthropic.
He said he is a firm supporter of AI agent workflows and often runs multiple agents simultaneously in his daily work, but has noticed a change worth watching: as AI models improve, obvious errors in their output become rarer, making it harder for people to remain vigilant.
In the past, people generally had to gather information, analyze it and make judgments before arriving at an answer. Today, that process may be compressed into a few seconds. AI generated content is often well structured, fluently expressed and correct most of the time, making users increasingly likely to skip independent thought and verification.
Osmani divides how people use AI into three levels, summarized by IT Home as follows:
First is “cognitive offloading,” in which a person delegates a task to AI but still checks the result carefully and takes responsibility for it;
Second is “deferral,” in which a person no longer treats AI output as a result requiring verification by default, but uses it directly as the basis for the next step;
Finally, there is “cognitive surrender,” in which users treat AI output as their own judgment and no longer think checking it is necessary.
In fact, the term “cognitive surrender” was originally proposed by University of Pennsylvania Wharton School researchers Steven Shaw and Gideon Nave. They discussed the phenomenon in a research paper titled “Thinking—Fast, Slow, and Artificial: How AI Is Reshaping Human Reasoning and the Rise of Cognitive Surrender.” The paper describes it as people accepting AI output without sufficient scrutiny and allowing AI’s judgments to replace their own intuition and thinking.
Osmani believes this shift usually happens gradually through repeated AI use, rather than all at once. A user might carefully read the first AI report, skim the second, and check only the outline of the third. Eventually, when someone asks about a specific conclusion, the user may find they cannot explain the reasoning behind it at all.
He links this phenomenon to “cognitive debt.” In software development, AI can generate large amounts of code and all the tests may pass, while team members’ understanding of the system’s architecture and implementation logic steadily diminishes. Only when the system needs to be changed one day do developers discover that no one truly understands how some of the code works.
Osmani does not, however, believe developers need to remember every detail of a large codebase. He notes that the Claude Code team also uses AI when needed to rebuild its understanding of particular code modules. What really matters is knowing what you do not understand and being able to understand it again, rather than knowing everything at all times.
To avoid “cognitive surrender,” Osmani recommends forming an initial hypothesis before asking AI a complex question, rather than relying entirely on AI for the answer. Even an inaccurate initial guess can help users retain a basis for independent judgment when they receive AI’s results. He also recommends actively looking for evidence that could disprove AI’s conclusions, instead of merely asking the model how confident it is in its answer.
In software development, Osmani believes the bottleneck in the AI era is shifting from writing code to verifying that it is correct. AI agents can already write code autonomously, run tests, read error messages and revise the code repeatedly until the tests pass. But the tests themselves may also be written by AI. If their logic contains the same error as the code, even a passing result on every test cannot prove that the software is free of problems.
He therefore emphasizes that developers should look beyond whether tests pass and be clear about exactly what the tests prove and which risks they do not cover. The acceptance criteria for work delivered by AI should be verifiable evidence, such as actual test results, runtime logs and screenshots, rather than a plausible sounding summary generated by AI. He also proposes that developers require agents to explicitly list the decisions they made independently, the assumptions they adopted and the conditions they ignored while carrying out a task.
As AI becomes able to carry out more and more tasks at the same time, another problem is emerging. Osmani calls it the “orchestration tax”: the gap between the speed at which AI generates work and the speed at which people can review it carefully and are willing to take responsibility for it.
When AI generated work keeps piling up and people do not have enough time to review it, adding more agents will not solve the problem. He believes teams should improve automated verification or slow the rate at which tasks are generated as appropriate.
Osmani also warns that extensive AI automation may lead people to gradually stop practicing their existing skills. He suggests that developers try writing some code without AI each week, revisit the skills they have stopped practicing each month, and decide which abilities remain worth retaining. Citing Anthropic research on AI use among its own engineers and researchers, he points to a “supervision paradox”: overseeing Claude requires programming ability, but excessive AI use may erode that ability.
Osmani stresses that this does not mean people should use AI less. It is fine to let AI handle tasks entirely when the cost of errors is low, but people should retain authority over key judgments in major decisions involving investments, career choices, product safety or effects on others.
He concluded: “You can delegate the work and most of the review to AI, but you cannot delegate the responsibility along with it.”