The Human Is the Loop
TL;DR
After a few weeks away from AI, a developer decides most of his agent use was a habit. The habit made him less curious and more anxious. From now on he wants to use AI on purpose. The human should be the loop, calling in an agent now and then instead of getting stuck inside the agent's loop.
The article
- He came back from a vacation without laptop or phone to eleven terminal tabs of paused agents. Unread Claude chats about taxes, landscaping ideas and policy docs were waiting too. He hadn't missed any of it.
- He says AI fed a belief he had never examined, that he should be able to do more. Old self-pressure turned into a stream of half-baked side projects that help no one. Those projects skip the learning, which is where the fun of tinkering used to be.
- Agents had become a buffer between him and stressful tasks. He spent hours talking ideas through with ChatGPT. He got back a sycophantic mirror whose output he never trusted for real work.
- Using AI as much as possible became a goal of its own, a "productivity ouroboros". He doubts agents are addictive the way algorithmic feeds are. He does see real dependency. Current AI valuations, he points out, rely on mass dependency.
- His alternative is deliberate use. For a work project he gave an agent a narrow task with as much context as he could. The task was searching messy systems he dislikes digging through. He spent his own time writing and reflecting.
Why it matters
Worth reading if you feel the pull to run many agents at once. The essay is short and candid, and it gets the guilt and compulsion across better than a summary can. Its closing example shows what a well-scoped agent task looks like.
The discussion
A warm thread. Many readers said the essay described them. A vocal group said it didn't and tried to work out why. The longest argument was over keeping AI out of writing and strategy. Next came attention and multitasking, with a long ADHD tangent. Another argument asked whether near-zero friction just means starting more things nobody needed.
Keep AI out of writing and strategy
The top comment said light AI touch-ups slowly pull your prose toward the LLM house style, and strategy advice fails without your full context. Against a claim that prose doesn't matter, replies said voice makes your message stand out, and an LLM knows more but not your context. Others said AI pads text rather than clarifying it, so you might as well send the prompt. Suggested fixes were reading human-written books and writing more without help, while a few let LLMs sanity-check a direction but not set it.
LLM's make your prose easier to *analyze* or *evaluate* but much harder to *hear*.
Do not fool yourself into thinking it has more context than you. It doesn’t. It has more knowledge. Those are very different.
Serial focus versus a dozen agents
The second big thread came from commenters with ADHD who keep one to four agent sessions on fixed jobs. They guessed that coping strategies built for a scattered mind carry over to managing agents. Commenters without ADHD also stick to one or two tabs, and called constant parallel agents the agentic version of believing more code means more output. One said heavy multitasking puts a non-ADHD brain into an "adhd-lite" state, and another worried that working serially will look outdated.
Perhaps it is the ADHD that actually helps me focus; that is, perhaps the coping mechanisms built for managing my countless disparate thoughts also work for managing multiple disparate agents?
People who keep dozens of agent plates spinning all the time must have a broken understanding of productivity where they think more code = better regardless if anyone needed the code.
Zero friction means saying yes to everything
With near-zero friction from idea to MVP, many said, impulses become projects before anyone asks whether they should. One commenter bought a domain for a meal-planner app before he stopped himself. Others said one big project, or the old discipline of saying no, matters more than ever, since the edge was always building the right thing. The common diagnosis was that LLMs create an illusion of free concurrency.
Success depends on focus and saying no to most things. AI makes it easy to say yes to everything. Instead of making my todo list shorter, it just makes me add more stuff to it.
LLMs provide you with the *illusion* of “free concurrency,” and it sucks people in.
Is the work worth doing, and by whom?
Some went further, saying if an LLM does a task well it may never have needed doing, and anything worth doing is worth doing yourself. Others objected that refusing faster delivery is like refusing email, and a junior developer coding by hand would be fired. For parents of young kids, lower activation energy is what makes a hobby project happen at all. Skeptics answered that speed doesn't cancel the maintenance burden, and worried about depending on tools they couldn't fix without an LLM.
With young kids, reducing the activation energy can be the difference between doing a project and never doing it at all.
Choose any 2: Cheap, fast, reliable. You might have a faster delivery time, but you cannot cheat on the maintenance burden.
Where AI pays off
Several listed where AI now pays off: research across scattered sources, reconciling accounts, smart-home commands, data checks, and learning practical skills like plumbing or cooking. Most drew the line between removing tedium and skipping thinking you needed to do yourself. One reader was confused by the ending, where the author says AI frees him to write. The author replied that he meant a deliberate split, where the agent gathers and reviews material while he reflects and writes.
There are tasks where AI genuinely removes tedious work and then there are tasks where the tedious or uncomfortable part is actually the thinking you needed to do
Consensus
Running many agents in parallel scatters attention and produces little of value. So does starting projects because AI makes them cheap. Focused, deliberate use works better, with people setting the direction and writing their own prose.
Dissent
A sizable minority said the essay didn't describe them. Some already work serially. Others said AI really does let them finish things, use fragmented time or learn new skills. A few argued that prose style doesn't matter when you're just delivering a message. Some said turning down faster delivery isn't realistic at work.