A few weeks away from AI reveals eleven paused agent sessions
A developer took a multi-week vacation, mostly away from his laptop and phone, and with it away from AI tools. Coming back, he found eleven cmux tabs open, each running multiple agents paused midway through various efforts, spanning professional and personal tasks. He also had unread badges on Claude chats covering everything from taxes to landscaping ideas to policy document review. He compares it to the old habit of hoarding browser tabs, but says it feels worse: each stalled session was something he had convinced himself he needed help with, and each one is now a small reminder of an unfinished task.
He traces the habit to using agents as a buffer between himself and stressful tasks, rather than as a tool for a specific job. One example: he would dictate his thoughts and talk them through with ChatGPT while driving or walking, telling himself it was productive brainstorming, while privately knowing the tool was acting as what he calls a sycophantic mirror. He never relied on the output of those sessions for real work, and what is left of the exercise is recordings and transcripts sitting on OpenAI's servers. A second example is trying to parallelize and automate tasks that no one had actually asked to have sped up, which he describes as a productivity ouroboros: using the tools mainly to maximize use of the tools.
He is careful to separate two claims. His hunch is that agent use is not intrinsically neurochemically addictive the way endless algorithmic feed scrolling is. But he does think there are real dependency and habituation effects once AI tools enter part of a person's workflow, and he notes that a large part of the tech industry is betting on widespread socioeconomic dependency on large language models, since current valuations depend on people becoming, in his words, hopelessly dependent on AI based tech.
His stated plan going forward is to be intentional about when to use AI and when to leave it out, and honest about what he actually gains or loses each time. As an example of the approach, before writing the essay he set up a single, narrowly scoped prompt in a tool called pi for a work project, giving it requirements, pointing it at codebases, wikis, schemas and past conversations, and asking it to come back with flagged issues and possible solutions, along with as much of his own context and open questions as he could supply. He does not expect the tool to hand him a finished answer or make him many times more productive, only to search across systems he is not good at navigating and possibly reveal gaps in his own understanding, freeing him to spend the rest of his time writing and reflecting. He ends on the line that gives the essay its title: the goal is not a human trapped inside the agent loop, but a human who is the loop, and calls the agent in only occasionally and thoughtfully.
Key facts
- After a multi-week break from AI, reopening his laptop showed eleven cmux tabs, each with multiple agents paused midway through various efforts.
- Unread badges spanned Claude chats on subjects from taxes to landscaping ideas to policy document review.
- He describes using ChatGPT while driving or walking as a sycophantic mirror, leaving behind rambling recordings and transcripts on OpenAI's servers rather than usable work.
- He calls one pattern a productivity ouroboros: using automation tools mainly to maximize his own use of the tools, for efficiency gains no one had asked for.
- His hunch is that agent use is not neurochemically addictive like endless feed scrolling, but he still sees real dependency and habituation effects, and ties current AI valuations to a bet on mass dependency.
Why it matters
The piece is a first person account of what heavy, low friction access to coding and chat agents does to daily work habits once the volume of open sessions outpaces what a person can actually finish or even track. It captures a specific and increasingly common failure mode: agents that are cheap to start and easy to leave running accumulate into a backlog of half finished, half remembered work, which is a different problem from the usual worry about AI output quality.
Who it affects
Anyone working heavily with concurrent AI agent sessions or terminal multiplexers for agents, and anyone who leans on a chat assistant as a running thinking partner throughout the day. The essay is written from inside professional AI-assisted development and knowledge work, not from a general consumer chatbot perspective.
How to use it
The author's own remedy is narrow scoping and explicit context rather than abandoning the tools: for a work project he wrote a single prompt for a tool called pi, told it which codebases, wikis, schemas and conversations to look at, supplied his own current thinking and open questions, and asked it to return flagged issues and possible solutions rather than a finished deliverable. His broader rule is to decide up front when AI use is warranted and when it is not, instead of defaulting to it.
How solid is it
This is a personal essay grounded entirely in the author's own experience and reflection, with no data, study or outside source cited. The account of eleven paused cmux tabs and the described habits are self reported and specific to one person's workflow; the broader claims about the tech industry's dependence on user dependency are presented explicitly as the author's own reasoning, not as sourced reporting.
Risks and caveats
The author is explicit that he is not claiming AI agents are neurochemically addictive the way feed scrolling is, and frames that as a hopeful hunch rather than a settled view. He does flag habituation and dependency as real effects, and raises, without citing any external source, the possibility that AI industry valuations are only justified if users become heavily dependent on the technology. The extracted text does not carry a byline or publication date, though the essay is hosted on the author's own site and carries both there; it does not specify what the flagged issues from the pi session actually were.
“We don’t want the human trapped in the agent loop. The human is the loop, and we tag the agent in occasionally, thoughtfully.”
— the essay's author, in its closing lines