Warning: Agentic Coding Workflows Risk Developer Obsolescence

The prevailing narrative in software engineering suggests that humans are evolving into mere orchestrators of artificial intelligence agents, a model known as agentic coding. Under this approach, developers generate high-level plans while AI handles the implementation, a shift often championed under the banner of Spec Driven Development. However, senior software engineer Lars Faye argues that this transition creates a dangerous trap, warning that the removal of direct coding friction leads to cognitive atrophy among engineers.

Faye contends that successfully managing these agents requires a high degree of critical thinking and architectural clarity to spot issues within thousands of lines of generated code. Yet, the very tooling designed to assist is proving to negatively impact the cognitive skills necessary for such supervision. A recent study by Anthropic highlighted this contradiction, identifying a "paradox of supervision" where the ability to effectively use coding agents depends on the coding skills that their overuse actively erodes.

Evidence of this phenomenon is already surfacing across the industry. Sandor Nyako, Director of Software Engineering at LinkedIn, reported observing skill atrophy within his team and has requested restrictions on using agents for tasks requiring critical thinking or problem-solving. Nyako emphasised that growth requires hardship and the development of mental muscle to think through problems, noting that without these skills, developers cannot question the accuracy of AI outputs.

Data supports these anecdotal observations, indicating a precipitous 47% drop-off in debugging skills among developers who aggressively incorporate AI into their workflows. Jeremy Howard, creator of fast.ai, has warned that outsourcing all thinking to computers guarantees obsolescence by halting the upskilling and learning processes essential for long-term competence. Howard argues that while tools like autocomplete and snippets have always aimed to reduce typing, the current trend of delegating the thinking process itself is fundamentally different.

Dax, creator of the OpenCode agent, noted that for many developers, the act of typing code is the process by which they figure out what to do. He contrasts this with the difficulty of sitting down to write a giant specification without implementation, suggesting that the current push toward high-level abstraction disconnects engineers from the technical realities of security, performance, and maintainability.

Beyond the loss of individual skill, there are significant financial and strategic risks associated with this dependency. Model providers are building on shifting sands, with costs and capabilities fluctuating rapidly. If an entire industry's skillset becomes dependent on specific vendor models, it could lead to a form of vendor lock-in where the expense of token consumption dictates the pace of work, leaving organisations vulnerable to sudden changes in pricing or model performance.

While Faye acknowledges that AI can be a powerful tool for learning and exploration when used responsibly, he advocates for a balanced approach where models act as secondary processes rather than primary drivers. The goal is to leverage productivity gains without sacrificing the tangible understanding that comes from engaging directly with the code, ensuring that engineers remain capable orchestrators rather than becoming obsolete.