I watched a video this week where one of the most-followed AI voices on LinkedIn proudly walked viewers through her setup: 36 proactive workflows, around 100 agents running while she sleeps. Impressive right? The comments were full of people asking how to build the same thing. Nobody seemed to be asking the more obvious question – “What am I really giving AI access to?”.
Here's the part the breathless productivity crowd tends to skim past...
For an AI agent to do useful work on your behalf, it needs access. Real access. Your inbox — yes, including the gynaecologist appointment confirmation, the slightly unhinged 1am message to your best friend, and the eleven unread newsletters from that dropshipping course you bought in 2019 and never finished. Your files, your calendar, your browser history (which we are not going to discuss). Sometimes your terminal and the ability to run commands on your machine.
Every one of those permissions is a door, and every door is one a bad actor can walk through if something goes wrong. Prompt injection (where instructions hidden inside an email or a webpage hijack your agent) is a known, documented, unsolved problem. You are, quite literally, handing a very eager intern the keys to everything and hoping nobody slips them a dodgy note.
And even setting the bad actors aside, do you actually want a language model reading the draft apology to your sister, the haemorrhoid cream receipt from Chemist Warehouse, or the screenshot folder labelled do not open, ever? Agents don't judge, but they do summarise, and they have no human filter. Somewhere in a log file, "user purchased one (1) tube of—" and I'll stop there.
I say this as someone who works in this stack every day and actually likes it. I'm writing this article with Claude's help, in fact, because I'm not too proud to admit a good tool saves me time. But there's a world of difference between using AI to sharpen your own thinking and handing it the wheel while you nap. One makes you better at your craft. The other makes you a middle manager for software that doesn't know what it doesn't know. Have you heard about how much AI lies when it doesn’t know something? They really have knowledge FOMO. Google “Gemini Lies”, there are a tonne of YouTube videos about it (we love a company that allows this kind of self-awareness).
Which brings me to the quieter problem nobody mentions in the 100-agents pitch. If you actually build that setup, congratulations — you are now a full-time approvals clerk. Your day becomes checking outputs, fixing hallucinations, re-prompting, reviewing drafts, untangling the thing that went sideways at 2am while you were asleep. The agents get the interesting work. You get the clipboard. Somewhere along the way the promise of "work while you sleep" became "never stop working, because the agents don't."
There's a reason a hand-made wooden spoon feels different. It carries the attention of the person who made it. You can't fake that, you can't scale it, and you certainly can't delegate it to something that doesn't know how wood feels. Attention is not the inefficiency in your workflow. It's the part of you that makes the work yours. You deserve to give something your whole mind. Your work deserves it too. A hundred agents working through the night will never add up to one thing made with care.
None of this means AI is bad, I'd be out of a job if I believed that! It means you should stop asking “how many agents can I run?”, and start asking “what am I actually trying to build, and does this tool make me better at it or just busier?”. For most people, a small number of well-understood tools used deliberately, will take you further than an army of autonomous ones you don't fully trust.
There's something worth noticing here. A single income used to be enough to run a household, until suddenly it wasn't. One brain used to be enough to do one job, until suddenly it won't be. Nobody announces these shifts. They just become the new floor, and everyone is expected to meet it.
Monique Farrell
11th April 2026