On "Agentic coding is a trap" post
Theo reacts to Lars Fay's article "Agentic coding is a trap". He agrees with most of it — cognitive debt is real, skills are atrophying, the gap between great and not-great engineers is growing. But he pushes back on the cost-going-up framing (cost-per-IQ-point is collapsing) and the vendor-lockin framing (skill issue, hop between providers). Core argument: AI is a slot machine that lets bad devs skip the pain of feeling dumb, which is exactly where learning happens.
Cognitive debt is the real concern, not tech debt
- AI is great at tech debt — brute force migrations, lint rule rollouts across 5000 files, the kind of work that used to require Google sheets and engineer assignment at Twitch.
- Cognitive debt is different: orchestrators drift away from the code, lose mental model, can't reason about new features. Simon Willison and Martin Fowler reporting this even with 30 years of experience.
- "If you weren't already experiencing a bit of it, you weren't shipping fast enough." Theo had this since pre-AI by virtue of shifting between many projects.
Costs: per-token up, per-intelligence down
- GPT-5.5 medium = GPT-5.4x high intelligence at <50% the cost. 5.5 low ≈ Sonnet 4.6 max at 1/8 the price.
- Bill goes up because usage goes up. Cost per IQ point is dropping ~8x in a few months.
- Companies hiring devs at $5M/yr who paid $100k/yr for Cursor last year and $500k this year — that pain is real, but it's volume not rate.
Slot machine and the pain of feeling dumb
- AI lowers the upfront cost of learning a thing. Pulling the lever hurts less than reading docs / debugging hopeless stuff.
- Skateboarding analogy: most quit before the ollie because feeling dumb hurts. In code you only had to feel dumb; now you can skip even that.
- "Your ability to say no as a developer has never mattered more. Your willingness to feel stupid is a superpower."
Who gets hurt
- Junior devs, career-transitioners, and "imposters who don't have imposter syndrome" — they never built fundamentals, AI now lets them stay there indefinitely.
- Austin Kennedy (22) tweet: 6-8 Claude Code terminals open all day, hits enter on 75% of responses, friends in their 20s all feel less sharp.
- Devs maintaining codebases bigger than any 5-year engineer would be comfortable with, learned to code a month ago, throw hands up when things break.
Debugging WITH AI is correct, debugging BECAUSE OF AI is broken
- Theo's outage on Ping: pasted error into Codex, ran three parallel agents with different theories, cross-referenced with error logs, fixed it. Would not have worked without knowing the system.
- "Anybody who says don't debug with AI is either stupid or bad faith. The problem is people leaning on AI because they don't understand the system."
The supervision paradox (Anthropic study)
- Using Claude effectively requires the skills that overusing Claude causes to atrophy.
- LinkedIn engineering director told 50 reports not to use AI for tasks requiring critical thinking or problem solving.
Coding = planning (theo method)
- Don't write a giant spec assuming you know how it works. Build the minimum viable shape first, learn from the friction, then write a good spec.
- Plan mode in agents isn't this. Pseudo-code back-and-forth with the model is closer.
- Cost of being wrong on folder structure / type system / data fetching has collapsed — willingness to experiment goes way up.
Disagreements with the article
- Vendor lockin: skill issue. T3 Code, hop between Claude / Codex / Cursor / open source. People stuck on Claude Code outages chose that.
- "If all your costs are fixed, you don't have a real business." Egg prices, RAM prices, compute prices — everything fluctuates.
The core unlock: knowing what code matters
- Most code wasn't worth writing if it ran once. AI flips this — now worth writing a 2000-line one-off NAS asset manager during a file transfer.
- For code that runs thousands of times a day: use AI to understand it deeply, find hot paths, refactor at scale.
- For one-off code: ship 10x more of it.
- When these two get confused, everything falls apart.
Closing
- Article quote (Jeremy Howard): "People who go all in on AI agents now are guaranteeing their own obsolescence."
- Theo: "If AI isn't making you smarter, you're using it wrong."