Anthropic engineer warns of cognitive surrender to AI agents

The rise of agentic coding

Addy Osmani works on the Claude Code team at Anthropic, a company heavily invested in agentic coding. This technology involves using AI agents to perform complex tasks, such as writing software, with less direct human intervention. As these tools become more common, the role of the human worker is shifting significantly.

Osmani noted in a recent essay that his daily routine no longer forces him to engage in much critical thinking. Because so much of his work is now handled by agents, the mental heavy lifting is being moved away from the human user. This trend is not unique to his role, as other engineers have expressed similar concerns about the shift.

Some engineers have described the transition as a move from being builders to being operators. Instead of creating the logic from scratch, they are managing the outputs of an automated system. While this increases efficiency, it raises questions about the long-term impact on human expertise and problem-solving abilities.

Defining cognitive surrender

Osmani categorizes how people interact with AI into three distinct behaviors. The first is cognitive offloading, which is the process of delegating a task to an AI while still maintaining ownership of the final product. He argues that this is a productive way to work and something that most people should do regularly.

The second behavior is deferral, which happens when the output provided by an AI becomes the primary input for the next step of a project. This creates a cycle where the human is following a path set by the machine. The third and most concerning behavior is cognitive surrender, a term coined by researchers at the Wharton School.

Cognitive surrender is defined as complete trust in the outputs of an AI without questioning the underlying logic. Osmani describes this behavior as lazy and warns that it creates what he calls cognitive debt. When a user can no longer explain the answer to a question, even by using the agent to help explain it, they have effectively ceded ownership of the work.

The risk of skill decay

A major concern for Osmani is the speed at which human skills can degrade when they are no longer exercised. He compares this to how people may have lost the ability to navigate using physical maps after phone apps took over, or how people may forget phone numbers because of contact lists. Coding is no different, and skills that are not practiced will likely disappear faster than people expect.

The goal for workers is not to stop using AI, but to ensure they are not losing the ability to think deeply about their work. Osmani suggests that the stakes of the task should determine how much a person delegates. In situations where the stakes are low, it is acceptable to delegate entirely to an agent to save time and energy.

However, for high-stakes work, humans must remain active participants in the process. Osmani emphasizes that the danger lies in the default behavior of the technology. Because AI makes it easy to stop thinking, users must intentionally build critical thinking into their daily routines to ensure they remain capable of independent work.

Verification as a habit

To combat cognitive surrender, Osmani proposes a framework for how engineers should interact with AI agents. He suggests that workers should adopt a practice he calls verification as a habit. This means actively checking the work produced by the AI rather than accepting it as the final result without scrutiny.

One way to do this is to ask the AI agent what it is capable of verifying and, just as importantly, what it cannot verify. By understanding the limitations of the agent, the human user can better identify where they need to step in. This keeps the human in the loop as a supervisor who understands the boundaries of the tool.

Osmani also suggests that workers ask themselves specific questions before and during their work. He recommends asking if they formulated a hypothesis before requesting help from an AI. He also suggests asking what skills they are no longer practicing. By staying aware of these questions, workers can maintain their expertise while still benefiting from the speed of AI agents.