Stop generating, start thinking

Agentic coding vs. prompt engineering

• Several commenters argue the author is “holding it wrong”: modern workflows use agents that index the repo, search the web, run tests, and iteratively refine code, making role-based, hand-crafted prompts largely obsolete.
• Others counter with concrete failures: agent+LLM confidently mis-advising resource management, producing segfaults, or generating incorrect API usage that a single manual web search would have avoided.
• Broad agreement that LLMs are not “thinking” but powerful heuristic engines guiding automated search; the surrounding tooling is doing much of the practical work.

Reliability and code quality

• Experiences diverge sharply: some say they barely hand-edit anymore and routinely one-shot tickets; others report verbose, poorly factored, badly integrated “slop” that increases review and maintenance costs.
• Tools appear strongest in mature, well-typed codebases with lots of examples and tests; weakest in greenfield projects, niche domains, or poorly documented libraries.
• Deep code review remains essential; critics doubt that genuinely scrutinizing every line can still be a net time-saver.

Productivity, backlog, and employment

• Proponents claim big productivity gains, enabling long-neglected backlog items and reframing developers as “assembly-line designers” and strategists.
• Skeptics note the absence (so far) of an obvious avalanche of valuable new software and worry that even if the tools work, they mainly accelerate job erosion and centralization of power.
• Debate over whether learning these tools now is essential future-proofing or a quickly obsoleted, shallow skill.

Understanding vs. outsourcing thinking

• Strong concern that heavy reliance on LLMs produces “prompt kiddies” who can modify behavior but never really learn the system, treating it as a black box.
• Others argue that focusing on observable behavior is acceptable and analogous to everyday reliance on complex infrastructure we don’t fully understand.
• Tension around “don’t commit code you don’t understand,” and what that means for training future developers if they seldom write code from scratch.

Ethics, data, and terminology

• Some emphasize that current LLMs are trained on unconsented human work and are deployed primarily to reduce labor’s economic power.
• Disagreement over the term “AI”: some reject it as misleading marketing; others argue “learning without intelligence” is incoherent and accuse critics of misunderstanding LLM internals.

Hype, metrics, and trajectory

• Dispute over whether we’re on the cusp of an agentic breakthrough or already seeing a plateau masked by hype.
• References to rising app counts and commit numbers are challenged as poor proxies for real value, and to the growing “garbogization” of software and the web.