Agentic Coding is a Trap

Lars Faye's case against the "spec-driven development with agents handling implementation" pitch. The core argument is a circularity: effective agent supervision needs the same critical-thinking and code-comprehension skills that heavy agent use measurably degrades. Faye lines this up with three other costs the industry framing tends to skip — skill atrophy across populations, vendor lock-in, and token-cost volatility.

The supervision paradox

The whole orchestrator-of-agents model depends on humans who can read thousands of lines of generated code and catch the bad parts. Faye points to documented evidence that those exact skills decay quickly under agent use — Anthropic's own research showed a 47% drop in debugging skill with aggressive workplace AI integration; LinkedIn's Director of Software Engineering told his team to avoid agentic tools for anything requiring critical thinking. Simon Willison, with nearly 30 years of experience, noted he can lose his "firm mental model of what applications can do and how they work." Junior developers get the worst of it: reviewing generated code teaches less than writing it.

This is the same trust-recursion 0xsid surfaces in agentic-coding-fatigue under the "verifier of the verifier" frame. Faye states it as a flat contradiction: the orchestrator role assumes a stable supply of competent orchestrators, but the role's daily practice undermines that supply.

Coding as thinking

Faye picks up the Dax (OpenCode creator) line — "me typing out code is the process by which I figure out what we should even be doing" — to argue that for many problems, code is the medium of thought, not its output. Replacing the typing with a spec just relocates the ambiguity: now the LLM fills the gaps with assumptions, you spend tokens on review cycles, and you finish further from the system than when you started.

This is the demand-side counterpart to Cantrill's "virtuous laziness": human time-pressure is what forces simplification, and the act of writing code is where that pressure operates. Skip the writing and you skip the constraint.

Inverted priorities

Pre-agent priorities (code comprehension, standards adherence, minimal readable code, reasonable turnaround) optimize for understanding. Agentic coding inverts the list — speed first, volume above quality — and Faye attributes the inversion to organizational incentives and token-usage metrics rather than to the tools themselves. The metric story is the same one ceo-ai-psychosis tracks (tokenmaxxing leaderboards) and that ai-great-leap-forward frames as backyard-furnace output worship.

Vendor dependency

Two costs the spec-driven pitch tends to elide:

• Outage exposure. Recent Claude outages left entire engineering teams unable to work — a sharp departure from "you need a keyboard and an editor." Subscriptions are now load-bearing.
• Token-cost volatility. Costs aren't fixed like salaries — they shift with each model release, and Faye cites reporting that new releases often need 2-3x the tokens to reach previous quality. Providers are subsidized, so the floor isn't where users assume.

This is the operational side of the subprime-ai-crisis / ai-subsidy-economics argument: at the team level, dependence on a subsidized supplier means the meter can change without you.

The responsible approach

Faye's prescription is not abstinence. It's a tighter envelope:

• Use LLMs for spec brainstorming, keep hands on implementation (20-100% manual depending on the task)
• Write pseudo-code to bridge intent and generated output
• Delegate only ad-hoc tasks and documentation lookups
• Never generate more code than you can review in one session
• Don't let agents implement patterns or domains you don't already understand

The "review in one session" rule is essentially the human-ceiling constraint from agentic-coding-fatigue turned into a shipping policy. The "don't delegate unfamiliar patterns" rule is the failure mode Maganti hit in his vibe-coding phase.

His own self-report: "I'm not going faster, but I'm doing better quality work." The Jeremy Howard line he closes on: "People who go all in on AI agents now are guaranteeing their obsolescence."

Where this fits

The wiki's anti-agentic-coding cluster now has four angles on the same observation:

• peril-of-laziness-lost — philosophical: LLMs lack the human time-constraint that drives simplicity
• no-silver-bullet-llms — theoretical: code generation isn't the bottleneck; essential difficulty is
• agentic-coding-fatigue — operational: the supervisor has a 4-5 hour daily ceiling
• This post — structural: the supervision role consumes the skills it requires; vendor lock-in and token volatility are real second-order costs

cult-of-vibe-coding is the procedural counterpart from the other direction — Cohen's audit-discuss-execute is the workflow Faye's "responsible approach" implies. ceo-ai-psychosis is the executive-level pathology the org-incentive part of Faye's argument is reacting to. clean-code-coding-agents supplies the mechanical reason agent-friendly code matters even if you accept Faye's full critique.

The new concept page skill-atrophy-supervision-paradox isolates the circularity itself, since multiple sources are converging on it.