Skip to content

Agentic Coding Is a Trap

Summary

A thesis arguing that full agentic coding workflows (Spec Driven Development, autonomous agent loops) create cognitive atrophy, skill degradation, and vendor lock-in — undermining the very capabilities developers need to be effective orchestrators. Notable for citing Anthropic's own research finding a 47% drop-off in debugging skills with aggressive AI use.

Details

The Core Argument

The industry is promoting "Spec Driven Development" — plan everything, let agents write all code, the human serves as a "good taste" reviewer. This creates four interconnected trade-offs:

Cognitive atrophy: The critical thinking and coding skills needed to evaluate agentic output are degraded by reliance on agentic output. It's a self-undermining loop: the more you use agents as your primary implementation tool, the less capable you become at reviewing what they produce. Anthropic's own study measured a 47% drop-off in debugging skills among developers who adopt aggressive AI assistance.

The skilled orchestrator problem: Evaluating thousands of lines of agent-generated code requires a developer who can think critically at the architectural level. But the path to that capability requires decades of hands-on coding friction — the exact friction that Spec Driven Development eliminates. Developers without that foundation are set up to fail as "orchestrators" because they lack the deep skill base needed to spot issues.

Junior developer damage: Junior devs learn through the friction of writing code, not reviewing it. Reviewing code is at best 50% of the learning process. Agentic workflows truncate this fundamental learning mechanism, compressing the path to expertise even as they raise the floor of the expertise needed.

Vendor lock-in: Full agentic workflows make teams dependent on vendor pricing and capability roadmaps. Token costs are unpredictable — you control employee salaries, but you have no idea what your monthly AI spend will be. New model releases follow "high benchmarks → hype → underperformance → nerf" cycles, creating perpetual cost instability. Sandor Nyako at LinkedIn (overseeing 50 engineers) restricted his team's agent use for this reason.

Faye's Alternative Workflow

Lars Faye advocates a middle path — not anti-agent, but anti-automation-first:

  1. LLMs generate specs and plans; human facilitates implementation
  2. Human still codes 20-100% of the time
  3. Write pseudo-code when interacting with models (closes distance between request and output)
  4. Use models as delegation utilities for ad-hoc generation and research
  5. Never generate more code than can be reviewed in a single sitting
  6. Never delegate work you couldn't do yourself

Metaphor: "Use them like the Ship's Computer, not Data" — assistance, not automation.

Why This Is Different From Past Tech Transitions

  • C++ → Java: didn't cause brain fog
  • Sysadmin → AWS: didn't eliminate networking understanding
  • Manager engineers losing coding edge: well-known phenomenon, but earned through natural career progression with decades of base skill

What's happening now is forcing developers who've never built that foundation into roles requiring that exact foundation. There's no 30-year path to skip.

Key Claims & Data Points

  • Anthropic's study found aggressive AI coding assistance produces a 47% drop-off in debugging skills — [source: raw/articles/larsfaye-agentic-coding-is-a-trap.md]
  • Sandor Nyako (LinkedIn SVP of Engineering) requested his 50 engineers stop using agentic workflows for certain work — [source: raw/articles/larsfaye-agentic-coding-is-a-trap.md]
  • Simon Willison reported not having a full day of coding — [source: raw/articles/larsfaye-agentic-coding-is-a-trap.md]
  • Jeremy Howard: "People who go all in on AI agents now are guaranteeing their obsolescence" — [source: raw/articles/larsfaye-agentic-coding-is-a-trap.md]
  • The token cost of agentic workflows is unpredictable and volatile, unlike fixed employee salaries — [source: raw/articles/larsfaye-agentic-coding-is-a-trap.md]

Open Questions

  • What empiric longitudinal research tracks coding skill retention for developers who shift to agent-mediated workflows?
  • Is there a quantifiable threshold of agent-usage where cognitive atrophy becomes irreversible?
  • How might developer education systems adapt to preserve coding fundamentals in an agentic-default era?

Sources

  • Agentic Coding is a Trap — Lars Faye (May 2026). Detailed thesis on the cognitive atrophy, skilled orchestrator problem, vendor lock-in, and junior developer impacts of Spec Driven Development.