After months of heavy reliance on Claude, Cursor, and Copilot, Scollon reports feeling more tired than before — and the fatigue feels qualitatively different from the tiredness of hand-writing code. He frames LLM-assisted development as a low-grade, continuous act of vigilance against plausible-looking but wrong output, which drains a different mental muscle than authorship does.
By submitting the essay and pulling 229 points, sosodev and the upvoting HN audience signaled broad agreement that AI-assisted coding produces a novel form of exhaustion. The comment section reportedly turned into a group therapy session, with staff engineers and CTOs echoing that continuous model-output review is neither writing nor reading but a third, more taxing mode.
The editorial argues that vendor telemetry from Anthropic, GitHub, and others measures only output metrics — lines of code, PRs merged, tickets closed — while ignoring what it takes out of a human to keep an LLM on the rails for eight hours. The honeymoon phase of agentic coding is ending, and the language for the hidden cognitive costs is only now being written down.
The editorial observes that senior engineers historically spent 20-30% of their day in review mode, but with Claude Code or Cursor's agent mode that number now easily hits 70-80%. This means engineers are functionally doing continuous code review on a collaborator who is fast and occasionally brilliant but structurally incapable of being trusted — a workload shift that explains the burnout pattern.
Alec Scollon's post 'I think I have LLM burnout' landed on the Hacker News front page this week and pulled 229 points — a solid signal for a personal essay with no framework, no benchmark, and no launch attached. The thesis is simple: after months of leaning hard on Claude and Cursor and Copilot, he's more tired than he used to be, not less, and the tired feels different from the tired he got from writing code by hand.
The comment section did what HN comment sections do when a piece hits a nerve: it turned into a group therapy session. Staff engineers, indie devs, and CTOs all showed up to say variants of the same thing. Reviewing model output all day is a distinct kind of cognitive load. It's not writing. It's not really reading. It's a low-grade, continuous act of vigilance against code that looks right and isn't.
This is the second or third wave of essays like this in the last few months — the honeymoon phase of agentic coding is quietly ending for a chunk of the practitioner audience, and the language for what replaces it is only now getting written down.
The productivity narrative around AI coding tools has been almost entirely output-side. Lines of code, PRs merged, tickets closed, features shipped. Anthropic and GitHub both publish numbers that look great in a keynote. What none of the vendor telemetry captures is the input-side cost: what it takes out of a human to keep an LLM on the rails for eight hours.
Writing code and reviewing code use overlapping but distinct parts of the brain, and the ratio has quietly inverted. A senior engineer used to spend maybe 20-30% of their day in review mode — PRs from teammates, their own diffs before pushing, the occasional debugging archaeology. With Claude Code or Cursor's agent mode in the loop, that number can easily hit 70-80%. You are, functionally, doing continuous code review on a colleague who is fast, confident, occasionally brilliant, and structurally incapable of knowing when they're wrong.
Scollon's essay points at something the benchmark charts miss: the failure mode of a hallucinated function isn't that it crashes — it's that it compiles, passes the test the model also wrote, and ships. The reviewer's job is no longer catching bugs; it's catching plausibility. That is genuinely harder work, and the industry hasn't priced it in.
The HN thread surfaced a second pattern worth naming: the loss of what one commenter called 'the walk to the whiteboard.' The small pauses where you'd stare at a wall, half-formed a solution, and let your subconscious close the gap. Those pauses used to be free. Now they cost a prompt, and the model fills them so fast that you never get the pause. Several commenters — including a few who are otherwise bullish on the tools — said their sense of ownership over the resulting code has degraded, and with it their memory of what they shipped last week.
None of this means the tools are bad. It means the workflow around them is immature. We spent forty years optimizing dev environments for the person who types the code, and we're now asking that same person to do a job the environment was never designed for.
A few things are worth trying, based on what actually seems to work for people in the thread who report being productive without being fried:
Rotate modes deliberately. Blocks of pure LLM-assisted work, then blocks of unassisted work, then blocks of review. Trying to do all three simultaneously — which is the default of any agent-mode IDE — is what produces the specific fatigue Scollon describes. A few teams have reported that a 'no-AI Wednesday' or a couple of unassisted hours in the morning genuinely helps, and not for nostalgia reasons.
Raise the bar on what you accept. If you wouldn't merge it from a mid-level engineer without questioning the approach, don't merge it from Claude either. The tools reward laziness because they're fast, and the exhaustion partly comes from the compounding cost of debugging code you accepted without understanding. Slower reviews now, less firefighting later.
Instrument the human, not just the model. If you're an engineering manager, the interesting metric isn't 'PRs closed per week' — that number goes up regardless. It's whether your seniors are still opinionated about the codebase in month six. If they've stopped pushing back on architectural choices, or they can't remember why a particular module looks the way it does, that's a real signal, not a vibe.
And if you're an IC: pay attention to whether you still enjoy this. Scollon's essay is honest that his enjoyment of programming has taken a hit, and a chunk of the HN thread agreed. That's not a productivity metric, but it's the one that predicts whether you're still doing this in five years.
The next generation of AI dev tools will probably have to solve for reviewer fatigue the same way the last generation solved for typing speed — because the bottleneck has moved and the tools haven't caught up. Expect to see more work on verification (Cursor's background agents, Anthropic's computer-use style verification loops, deterministic linting layered on top of model output), better UI for showing the diff between what you asked for and what you got, and eventually some honest research on what an 8-hour day of LLM-mediated coding does to a senior engineer's brain. The essay went to the front page because a lot of people recognized themselves in it. That recognition is the leading indicator; the tooling will follow.
From my experience, there are mainly 3 burnout reasons. 1. Multi-tasking is the top one. I usually have to frequently switch between 3 to 5 agent windows which are on different things. It's extremely exhausting when each round takes a few minutes. Before coding agent era, I believe most develop
I've started feeling slightly physically ill when I read Opus output for hours straight. This article rings very true for me. I've started complaining about it with my team; at least have a personal style guide in your agent rules that eliminates emdashes, the "it's not X, it
I am feeling very tired. Since I’ve started working with LLMs my output as a solo dev has easily gone up 20x. I’m closing client projects including ones that previously would have been far too ambitious to take on alone. Long running codebases are getting features that have dragged out for months or
I’m currently doing a project with someone who only uses LLMs, and it’s exhausting and mentally draining.Whenever I give feedback on something, the answer is just “let me tell Claude”. The person has no understanding of how everything works, and most of the code reflects that.The other day he hardco
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I don't think I have a "burnout", but LLMs are really exhausting due to amount of pressure they generate. No one is really pushing me to increase my workload, but at every moment there is always something ready, done by my clankers or clankers of other people that I could be unblockin