Agent Orchestration Requires the Skills It Atrophies

·https://larsfaye.com/articles/agentic-coding-is-a-trap
articleaiagentic-codingsoftware-developmentagent-loopscognitive-debtvendor-lock-in

pushes joelclaw agent loops toward bounded delegation, reviewable diffs, and human-owned code comprehension instead of unattended code churn

Lars Faye is pushing back on the fantasy that AI does the coding and the human just floats above it all as the tasteful little orchestration wizard. His point is sharper than “don’t use agents.” It’s that supervising coding agents requires the exact coding judgment that agent overuse can erode.

That matters because the popular workflow adds distance: write a spec, pull the slot machine lever, review a mountain of generated code, repeat. Faye names the tradeoffs directly: more surrounding system complexity, skill atrophy, vendor lock-in, and token costs that move under your feet. The Anthropic line he cites is the nasty little hinge: “supervising Claude requires the very coding skills that may atrophy from AI overuse.”

The useful move is his inversion: demote the model. Use LLMs for specs, planning, research, docs, small delegation, and interactive clarity, while keeping the human close enough to the code to understand the shape of the thing. Dax Raad, creator of OpenCode, says the quiet part well in the linked interview: typing code is part of how he figures out what the feature should even be.

For joelclaw, this is a guardrail against agent-loop bullshit. Inngest, Restate, and Pi can make autonomous work durable, but durability doesn’t make generated work comprehensible. The sharper pattern is bounded delegation with small reviewable diffs, explicit tests, and human-owned understanding.

Key Ideas

  • Faye frames “agentic coding” as a supervision paradox: effective use of coding agents depends on coding skill, while overuse can weaken that skill.
  • The article argues that LLMs accelerate the wrong default metric: code generated per unit time, not understanding, concision, maintainability, or review quality.
  • “Coding === Planning” is the strongest operational idea: writing code, types, folders, and rough function boundaries is often the thinking process, not disposable typing labor.
  • Anthropic reported a debugging-skill drop in its coding-assistance research, which Faye uses as evidence that this is not just nostalgia from cranky senior engineers.
  • Vendor lock-in here is not just dependence on Claude, OpenAI, or another model provider. It is dependence on paid token access for work that used to live in a developer’s own problem-solving muscles.
  • The practical counter-pattern is “use them like the Ship’s Computer, not Data”: delegate bounded tasks to LLMs, but keep the developer actively engaged in implementation and review.

Links

← All discoveries