An Anthropic engineer has raised concerns that leaning too heavily on AI agents could lead to what he calls cognitive surrender, where professionals stop exercising the critical thinking that underpins their work. Addy Osmani, a software developer at the company, shared his reflections in a post that highlights a growing unease about skill atrophy in an era of increasingly capable AI coding tools.
Osmani’s central worry is not that AI agents are inherently harmful, but that their convenience quietly removes the everyday friction that once kept human judgment sharp. When much of a developer’s work runs through automated agents, he suggests, nothing in the daily routine forces the kind of deliberate reasoning that complex problems used to demand. Skills that go unpracticed, he notes, tend to fade faster than people expect.
The erosion of critical thinking
The engineer described a pattern that may feel familiar to anyone who has used AI for extended periods. A user reads the first AI-generated answer closely, then skims the next one, and merely glances at the structure of the one after that. By the time someone asks about a specific line, the user cannot explain it, even with the agent’s assistance. Osmani frames this as an accumulation of cognitive debt, a slow build-up of unexamined output that leaves the human less capable of independent analysis.
Cognitive offloading to AI is not inherently wrong, according to Osmani, as long as people retain ownership of the work. The danger arises when AI output is accepted without checks or rigorous examination. At that point, the human is no longer directing the process but merely receiving it, and the ability to spot errors or question assumptions begins to erode.
Responsibility still rests with people
Osmani also pushed back against the argument that losing certain skills is acceptable because AI can perform those tasks better anyway. Regardless of how much assistance AI provides, he insists, a human must still take responsibility for the outcome. That principle of accountability means someone needs to understand what the AI has done and why, even if the machine handled the mechanics.
For Osmani, the answer is not to abandon AI coding agents but to use them deliberately. He offered several suggestions for avoiding cognitive surrender:
- Practice core skills regularly rather than letting them lapse.
- Formulate ideas before turning to AI for answers.
- Pay close attention to what the AI generates instead of skimming.
- Make sure you understand the decisions the AI has made.
These practices are meant to preserve the human role as the responsible party, not just the operator of a tool. The distinction matters because AI agents can produce plausible results that hide flawed reasoning, and only an engaged human is positioned to catch those flaws.
Choosing what to let go
Osmani acknowledged that some skills will inevitably be lost as AI systems improve. The key, he argues, is to be deliberate about which skills are allowed to fade and which must be protected. Letting them disappear without thought is the real risk. By deciding consciously what to maintain, developers and other professionals can keep the capabilities that matter most while still benefiting from AI’s assistance.
The warning arrives as AI coding agents become more deeply embedded in software development workflows, handling everything from generating code to debugging and documentation. As those tools grow more capable, the temptation to delegate more of the thinking grows alongside them. Osmani’s message is a reminder that the human capacity for judgment is not a fixed resource. It needs exercise, attention, and a willingness to stay engaged even when the machine offers an easy way out.
For anyone who regularly asks an AI for help, the question worth considering is whether the assistance is supplementing their skills or quietly replacing them. The difference may not be obvious in any single interaction, but over time it shapes who remains capable of understanding the work and who merely watches it happen.
