# Agentic Coding is a Trap

# Agentic Coding is a Trap

**By Lars Faye**

## Remaining vigilant about cognitive debt and atrophy.

The current industry narrative suggests that traditional coding is obsolete and "Spec Driven Development" represents the future. In this model, developers create detailed plans while AI agents handle implementation, positioning humans as orchestrators who review outputs and steer the process. However, this approach introduces significant trade-offs worth examining.

### Key Trade-offs of Agentic Coding

- Increased complexity in surrounding systems to manage AI's non-determinism
- Skill atrophy across developer populations
- Vendor lock-in for individuals and teams
- Unpredictable and rising token consumption costs

The approach depends critically on skilled developers who can identify issues in thousands of lines of generated code. Yet this is precisely where the contradiction emerges: the very critical thinking skills necessary to effectively supervise AI coding are being undermined by reliance on these tools.

### Different from Previous Tech Transitions

Unlike earlier programming language shifts, the cognitive impacts are already measurable and significant. Developers with decades of experience report losing their mental models of their own applications. Junior developers face particular challenges, as reviewing generated code provides only partial learning compared to writing code directly.

As Simon Willison noted, even with nearly 30 years of experience, developers can lose their ability to maintain a "firm mental model of what applications can do and how they work."

### The Supervision Paradox

Anthropic's research identifies a critical problem: effectively using Claude requires supervision, yet supervising Claude requires the coding skills that may deteriorate from overuse. LinkedIn's Director of Software Engineering observed this phenomenon across his team, requesting engineers avoid agentic tools for "tasks that require critical thinking or problem-solving."

Skill degradation occurs rapidly—within months in documented cases—creating a circular problem where orchestrators lose the expertise needed to orchestrate effectively.

### Inverted Priorities

Before AI, developer priorities typically emphasized:
- Code comprehension and codebase alignment
- Adherence to standards
- Minimal, readable code
- Reasonable turnaround time

Agentic coding inverts this list, prioritizing speed above understanding and conciseness. While these tools *could* be used thoughtfully, organizational incentives and token-usage metrics drive toward volume over quality.

### Coding as Thinking

Many developers think through problems *by* writing code. This isn't mere drudgery—it forces consideration of security, performance, maintainability, and user experience. Dax, creator of OpenCode, explained: "When working on something new or challenging, me typing out code is the process by which I figure out what we should even be doing."

Replacing this with specifications introduces ambiguity that LLMs fill with assumptions, creating more review cycles, more token consumption, and greater disconnection from the final product.

### Vendor Dependency

During recent Claude outages, entire engineering teams found themselves unable to work. Developers had already become dependent on subscription access to continue their jobs—a dramatic shift from needing only keyboard and editor.

Token costs remain unpredictable and volatile. Unlike fixed employee salaries, token consumption fluctuates with each model release. The Verge reported that model providers are heavily subsidized, and each new release often requires 2-3x more tokens to achieve previous results.

Anthropic's own research showed a 47% drop in debugging skills with aggressive AI integration in workplaces.

### A Responsible Approach

Rather than fully delegating to agents, developers can:

- Use LLMs for brainstorming specs while maintaining hands-on implementation (20-100% manual coding depending on task)
- Write pseudo-code to bridge the gap between requests and generated output
- Delegate only ad-hoc tasks and documentation research
- Never generate more code than can be thoroughly reviewed in one session
- Avoid having agents implement unfamiliar patterns or tasks

This "secondary process" approach preserves understanding while gaining productivity benefits. As the author states: "I'm not going faster, but I'm doing better quality work."

### Historical Caution

Jeremy Howard, creator of fast.ai, warns: "People who go all in on AI agents now are guaranteeing their obsolescence. If you outsource all your thinking to computers, you stop upskilling, learning, and becoming more competent."

The parallels to social media adoption—where long-term implications were underestimated—suggest caution. This time, the stakes involve the competency and autonomy of an entire profession.

The fundamental risk remains: you cannot understand code without engaging with it, and disengagement leads to skill loss that undermines the very supervision these tools require.