I Hired an AI to Do My Chores. Now I Maintain the AI.

AI agents promise to automate away the tediousness of modern life—the overbilled rental cars, the iCloud storage alerts, the changing of leaked passwords. Jack Cheng, Every’s senior editor, put that promise to the test. But instead of his AI agent maintaining his digital life, he ended up maintaining his AI agent. From there Jack explores Stewart Brand’s philosophy of “nested maintenance,” COBOL Cowboys, and civic technologists watching Claude Code attempt to modernize government benefits systems. Read on for an account of what it means to hand our most tedious obligations to machines—and what we only come to understand about broken systems by struggling with them ourselves.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

I set up OpenClaw in hopes that it would automate away the petty bureaucracy of modern life.

Maybe a Claw could keep my iCloud storage account from constantly hitting its cap, or go through my over 1,000 different online accounts and change all the passwords that were leaked by hackers onto the dark web. Maybe, I thought, it could even help my family sort out a medical bill we got from an unexpected hospital visit while traveling before we switched from our old health insurance to our current one.

It’s not the first time I’ve tried to tackle this problem. Since 2023, I’ve been hosting what I call a “Digital Mending Circle.” With a small group on Zoom, I tend to the maintenance tasks that accrue around a digital existence. Instead of darning socks or patching jeans, we update personal bios, organize photos, file expense reports, or even just catch up on email.

These activities can feel surprisingly daunting, given how trivial they are in the grander scheme. They involve re-familiarizing yourself with systems you only use occasionally (where’s that page in my WordPress admin panel again?) or facing clean-ups—the 571 items on your desktop, the gigabytes of blurry and duplicate photos across multiple apps—that will just need re-tidying months later. Maybe that’s why we so often neglect them.

Now, generalist AI tools like Claude Cowork and specialist tools like Sparkle, Every’s AI file organizer, can do many of these tasks for you—and swiftly. They’re chores that Claws, or whatever forms personal AI agents take in the future, could do for you without your ongoing input.

So I’ve been pondering this question: What does maintenance look like when you have AI running your digital life?

Auto-scale AI workloads with ease

Traditional databases struggle with unstructured data, slowing AI workflows. MongoDB natively handles JSON-like, unstructured data, giving you freedom to store, query, and scale diverse datasets—whether text, images, or sensor data—easily and flexibly, without rigid schemas.

Try MongoDB Atlas free

Want to sponsor Every? Click here.

The maintenance of everything

“Maintenance is absolutely necessary and maintenance is optional,” says Whole Earth Catalog and Long Now Foundation founder Stewart Brand in Maintenance: Of Everything, Part One. Optional because we can put it off in the moment, necessary because putting it off for too long can lead to disaster.

We learn, in the book, of boats whose maintainability resulted in very different outcomes for three sailors competing to first circumnavigate the globe. We discover how maintenance attitudes in militaries can sway entire wars. Good and poor maintenance can both have profound consequences.

Maintenance is virtuous. But it’s also rarely seen as heroic. If it were, maybe we wouldn’t be so bad at it. Various explanations exist for why we deprioritize maintenance, ranging from cultural values (we prize new invention over care for the existing), psychology and economics (we discount what isn’t immediately gratifying), and social class (we associate many maintenance jobs with minimum-wage work done by marginalized workers).

For me, many of these attitudes are embodied in Pixar’s 2008 animated film about a solitary garbage robot, Wall-E. The cheerfulness with which Wall-E performs his Sisyphean task of collecting and compacting tiny robot-sized cubes of trash makes us care for him. But he doesn’t become a hero until he leaves behind his duties to follow his love interest across the galaxy. At the end, we learn his quest is part of a larger story of failed maintenance—of the earth and its natural systems.

“Nearly everything worth maintaining is nested,” writes Brand, “in something larger, even more worth maintaining.”

Admin nights

For over six years, journalist and author Chris Colin has been hosting in-person gatherings akin to my digital ones. At these “Admin Nights,” Colin and friends gather with their laptops to cancel streaming service subscriptions, file insurance pre-authorization forms, dispute erroneous credit card charges, and, more generally, try to pull themselves out of the morass of maddening tasks that swallow modern life.

In the process, they’ve grown more aware of the sources of that madness, like the rise of subscription models and the breakdown of unions, regulators, and community groups that once shielded us from consumer abuse. Will AI eliminate this administrative friction, or only worsen it?

Colin does not dismiss the possibility that AI could make a meaningful difference, but said the group has developed a jaundiced long view of new technology. “I came to San Francisco in 1998, and I’ve seen a lot of things come through that promised to be amazing and weren’t,” Colin said to me over the phone. “Or things that promised to be amazing and were, and then were discontinued or evolved or broke after a few years.”

Even if AI ends up being “amazing,” I also wonder if smoothing over the friction would obscure its underlying causes, like the convenience-laden hoverchairs in Wall-E that numb senses and distract the humans from their world’s ills. Maybe it’s through the act of collectively dealing with that friction that Colin and his Admin Night group start to understand—and perhaps change—the broader systems to which most of us belong.

The COBOL cowboys

Those broader systems are facing a maintenance crisis of their own. Consider COBOL, a programming language designed in 1959, that handles transactions in 43 percent of all banking systems worldwide. It runs primarily on IBM mainframe computers, and still underpins much of the infrastructure behind public benefits like unemployment and food assistance that millions depend on.

Institutions haven’t modernized their COBOL systems because the process is a pricey, years-long affair. Meanwhile, the number of people in the world who have the expertise to maintain these systems shrinks by the month. When the situation is dire, organizations often call on the COBOL Cowboys, run by 82-year-old cofounder and CEO Bill Hinshaw and staffed with some 600 COBOL engineers, many of whom got their start in the 1960s and ’70s programming those original mainframes. Heroes, for sure, but maybe only because they come in when these systems break.

So when Anthropic announced last month that Claude Code can now automate the most time- and cost-intensive aspects of the COBOL modernization process, IBM’s stock fell by 13 percent, its single largest drop in 26 years.

Not great for IBM’s business, but potentially transformative for people stuck on the other side of these systems. Adam Selzer, co-founder and senior director of Detroit non-profit Civilla, which helps local and state governments redesign how they administer benefits to their constituents, says that legacy technology is a significant barrier to change within these institutions.

“[It] could be one of the most significant developments in civic technology in this decade,” says Selzer of the Anthropic news. “Or could lead to it.”

What’s Claw got to do with it

I personally haven’t gotten as far as overhauling the economy’s financial infrastructure. I’m still trying to contact my bank about an overbilled rental car and make my “iCloud Storage Full” alert go away.

So at a recent Digital Mending Circle, on the promise of a more frictionless life, I set up OpenClaw. Or started to set up is more accurate. It took several hours beyond my allotted life maintenance session for me to get my Claw, Pip, working. Once Pip was up and running, I had it send my partner and me morning briefings of our day’s schedule and childcare duties, and prepare budget summaries that we could go over in our weekly household meeting—augmenting a system we already had in place to help us stay afloat.

But sometimes Pip would format the budget summary in a weird way or forget how to access our transactions. Sometimes the morning briefings wouldn’t come—a result of my Claude Code account hitting its monthly limit or Pip incorrectly updating a setting in its own configuration file. At one point, two weeks’ worth of Pip’s memories simply vanished. In each of these cases, I would have to then go in and figure out why Pip broke, and how to fix it.

Instead of Pip helping maintain my digital life, I was maintaining Pip.

But in the process, I’ve become intimately familiar with the OpenClaw configuration file. I’ve come to better understand the cascade of markdown instructions that Pip reads for context, the art of managing Pip’s short- and long-term memory, and the way it uses skills and tools to do its work.

I’ve also come to learn, true to Brand’s words and Colin’s Admin Night group, the wider context around Pip: the ecology of frontier models, AI agents, new and existing software, and the people trying to use and build these tools. And once I could see this more clearly, I could start to imagine these systems working and interacting in a different way.

Then again, I still have to look into that rental car bill.

Jack Cheng is a senior editor at Every. He is a creative generalist and the author of critically acclaimed fiction for young readers. He writes an occasional newsletter called Sunday.

To read more essays like this, subscribe to Every, and follow us on X at @every and on LinkedIn.

We build AI tools for readers like you. Write brilliantly with Spiral. Organize files automatically with Sparkle. Deliver yourself from email with Cora. Dictate effortlessly with Monologue.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

For sponsorship opportunities, reach out to [email protected].