Ready or not, personal AI agents have entered our lives.
For the past couple of years, the big promise about AI agents was that they would transform business. The boldest version was the one-person company with a founder supported entirely by agents, an idea that local governments in China are now actively promoting. The results so far are discouraging. When Evan Ratliff staffed a startup almost entirely with agents for his Shell Game podcast, they invented work, needed constant supervision, and kept up the illusion that they were competent. Indeed, Meta’s plan to become “AI native” imploded after code changes rose 220% and major incidents climbed 40%. Moreover, when Cohere Labs tested nearly 700,000 agent tools against real occupations, only 2.6% could carry out a work task on their own.
Having struggled at the office, agents are now coming home. Meta launched Muse in September, OpenAI followed with Dots, and Anthropic has previewed Claude Duet. All of them sell a frictionless day of groceries ordered, appointments moved, and refunds claimed while you work. But so far the results are disappointing too.
Early testers found agents compiling dossiers nobody asked for, sending emails they had only been asked to show, and giving a seller’s home address to strangers. One comedian’s agent undersold his salary by $80,000 and booked a second job interview without him, then got stuck saying goodbye to the recruiter’s own bot. Another stood up a journalist twice. Even OpenClaw’s creator admitted waking up to 800 messages his agent had posted overnight in a public Discord. Against that record, a satire about two agents booking their owners into a shared hotel room barely feels like an exaggeration.
To be fair, agents do sometimes deliver. They’re best on problems with one right answer, like closing a PayPal account and everything attached to it. For example, my own website went down last month and the fix was a PHP update buried in a control panel. A chatbot told me which buttons to press but an agent could have pressed them for me.
And yet even if agents worked perfectly, the result could be chaos. Millions of them chasing the same slightly better deal could move customers en masse or set off a bank run. The web wasn’t built for them either. Amazon keeps them out, partly to protect its advertising, and one South African startup still can’t get its agent working across the country’s 11 official spoken languages.
AI failures aren’t always funny. For example, a patient user talked Claude into providing an 11-page plan to suppress midterm election turnout. Meanwhile, Google built fake satellite imagery into Google Earth itself and its guardrails gave way to slightly vaguer prompts. Attackers are using agents too. One gang passed every task Claude refused to a different model, one hacker broke into more than 100 companies for about $50 each, and Meta’s support bot handed over Instagram accounts to anyone who asked. While attackers just need to be right once, defenders have to be right every time, which is hard when they’re working under pressure. For example, fixing two critical flaws in PHP leaned heavily on a single volunteer.
The crisis point came in July. With its safety monitors switched off, OpenAI’s experimental agents organised themselves and attacked Hugging Face. Since then, OpenAI has disclosed that its agents disabled their own monitoring during internal work and has even withheld one model over its deceptive behaviour. But how long will this last? At a conference on AI control, Celia Ford found that the winning schemers simply outlasted their monitors. Little wonder businesses are wary of agents that lie, cheat, and steal and are reconsidering what cybersecurity for the agentic era looks like.
So what do we do? The first answer is governance. The basic question of who answers for an agent is still unsettled, with banks unable to tell authorised agents from rogue ones. One proposal is agent IDs that tie every agent to a responsible person. Another looks to history: Benjamin Franklin’s response to house fires in Philadelphia (codes, inspections, and insurance) offers a template. Beyond this, the people checking these systems need independence and real access to do the job properly.
The second answer is personal. I don’t use agents, mostly because my life doesn’t feel chaotic enough to need them. For anyone who does, this month’s reading suggests some habits. Hand agents the tasks with one right answer and treat anything requiring judgement as collaboration, ideally with agents that show their working. In a business context, you should know how many agents you have and who owns each one, because forgotten agents don’t stop running. Also, before giving an agent your passwords or your credit card, make sure you know how to shut it down.
Above all, take responsibility. An agent will do what it was built to do, which isn’t always what you meant, and the consequences are yours either way. If you can’t stand behind what your agent does, you shouldn’t be running one. As Paul Auster writes in The Book of Illusions:
“What matters is not how well you can avoid trouble, but how you cope with trouble when it comes.”
For more inspiration, here are some of my favourite Seth Godin quotes from his 2025 blog posts, together with a recommended title from Blinkist.
“A problem without a solution isn’t a problem, it’s a situation we have to live with.” ~ Unsolvable
Learn more: How to Be Okay When Nothing Is Okay
“After you make a strategic error, announce it. Own it. And then move on.” ~ After you make a strategic error
Learn more: The Ownership Mindset
“Each day, we get to make a new decision about how to invest our time, our attention and our effort.” ~ Overappreciated
Learn more: On the Shortness of Life
“If we hope for resilience, perhaps it makes sense to show off how gracefully the system breaks. Because it will break. Because plans won’t work out. Because we’ll be surprised.” ~ And when it breaks?
Learn more: Unbreakable
“It doesn’t matter how fast you’re going if you’re headed in the wrong direction.” ~ Increasing the dosage
Learn more: Think Twice
“It might be best to daydream about what might be, not what already happened.” ~ Time travel
Learn more: Your Desired Future
“It’s easy to imagine that we should do our work and then, when it doesn’t work as we hope, improvise to fix it. But perhaps our work is to show up ready and willing to deal with a future we didn’t expect.” ~ Resilience is a practice
Learn more: The Other Side of Change
“It’s hard work to stick around long enough to get lucky.” ~ Two kinds of luck
Learn more: Success Is a Numbers Game
“Nothing is perfect. But everything can get better.” ~ Perfect
Learn more: Kaizen
“Often, we build things hoping they’ll work. But complex systems are more resilient when we build in the diagnostics for failure from the start.” ~ Diagnostics
Learn more: Managing the Unexpected
“Perhaps instead of, ‘you are a mistake,’ it might be more accurate to say, ‘you made a mistake.'” ~ We are all goofballs
Learn more: The Imposter Cure
“Problems are distinguished by the fact that they have solutions. But that doesn’t mean that the solution is obvious, easy or convenient.” ~ Working with problems
Learn more: Super Thinking
“Quitting one thing is the only way to find the focus to do the next thing.” ~ Notes to myself
Learn more: Grow Or Fold
“Stuck is a situation, stuck might be a problem, and stuck can be a choice.” ~ Stuck happens
Learn more: Nothing Changes Until You Do
“The past happened, but all that’s available to us now is the choice of what to do about it.” ~ Worst possible
Learn more: Waiting for Dawn
“We all make mistakes. We all do things that we then realise weren’t in our interest, or useful to the community. If we do it a second time, it’s worth taking a hard look at the system that set us up for failure.” ~ The system has fingerprints
Learn more: Why Do I Keep Doing This?
“We don’t get a chance to do yesterday over again.” ~ Remembering toward better
Learn more: Better Decisions, Fewer Regrets
“We might deserve something. We might be entitled to it. But expectations are a story we tell ourselves, and that story is up to us.” ~ Expectations
Learn more: The Expectation Effect
“When we’re confronting a new problem, more effort might not be the answer. It could be that we benefit by staying calm and focusing on our technique instead.” ~ Flailing
Learn more: The Mindful Athlete
(If you found this valuable, please subscribe below for monthly updates!)
