An Anthropic engineer warns about 'cognitive surrender' as AI agents take over his work

• Anthropic engineer Addy Osmani warned about "cognitive surrender" among coders.
• "Nothing in my day-to-day forces critical thinking on me," he wrote, citing AI agents as the culprit.
• Osmani advocated for picking selectively where to offload — and where to ask follow-up questions.

Addy Osmani is thinking hard about thinking.

Osmani works on the Claude Code team at Anthropic. The company is all in on agentic coding — he helps create its AI coding product, after all — and has reduced human code writing.

"Increasingly, nothing in my day-to-day forces critical thinking on me now that so much of my work runs through agents," Osmani wrote in a Tuesday X essay.

Engineers have expressed similar concerns in the last year: AI has turned them more into operators than builders. One OpenAI alum told Business Insider that AI coding was tedious and boring.

Osmani didn't want engineers to stop using AI; he wanted them to learn to think more deeply.

The risk of "cognitive surrender"

Osmani laid out his thesis: "You have to build critical thinking into your day, because by default nothing will do it for you."

There are three types of diminished thinking with AI use, he wrote. First is "cognitive offloading": the process of delegating tasks to AI while retaining ownership of the products. "I do this all day," he wrote. "You should, too."

The second is "deferral," where AI's output becomes your new input. The third is "cognitive surrender," a term coined by two researchers at the Wharton School. This is complete trust in AI outputs, which Osmani described as lazy. Surrendering too often creates "cognitive debt."

It's not yet clear whether AI is making us dumber or changing the way we think. Researchers Business Insider spoke with saw some bad signs, but cautioned against jumping to conclusions.

As for Osmani, he often sees this surrender of critical thinking. It's not the worst thing, he wrote; it's merely a question of the task's importance.

"Where the stakes are low, feel free to delegate entirely," he wrote. "When you can't explain the answer to a question, even by way of your agent, you've ceded ownership."

How to think with AI agents

Osmani shared a list of questions that he asks himself. They include, "Did I formulate a hypothesis before asking?" and "What am I no longer practicing?"

Again, he's not asking engineers to think hard about everything, or even most things. In response to a question differentiating help versus offloading, Osmani wrote: "Both are OK, as long as you know which one you're doing."

Engineers should also build loops to ensure the code that AI has written works, Osmani wrote. He called it "verification as a habit" and told engineers to ask the AI agent what it could verify (and what it couldn't).

Skills often degrade over time. How many of us can navigate using a physical map after phone apps took over? How many of us remember our friends' phone numbers when we can merely save them to contacts? Coding is no different.

"Skills that you stop exercising are going to go away, and probably faster than you might guess," Osmani wrote. "Keep doing the ones you're not willing to let go."