A small experiment with a large implication: what an AI chooses to remove says more about us than about the machine.
I had, of course, read the stories about OpenClaw. Installing an AI agent at home had started to feel less like curiosity and more like a calculated risk. Bank accounts emptied. Entire inboxes erased by an Agent that had taken its instructions too literally. People who had casually asked to remove “irrelevant emails” and discovered that relevance is a dangerous thing to outsource.
That is where the problem stops being technical and becomes existential. An Agent does not hesitate. It executes. And if we are honest, brutally honest, how much of what sits in the average inbox deserves to survive? Mine is mostly a graveyard of postponed responsibility. Messages I do not want to answer, left in plain sight on purpose. Not archived, because they need to remain visible. A quiet pressure. A promise to a future version of myself that will likely never materialize.
An Agent trained on human behavior understands this perfectly, and then improves on it. It does not share our hesitation. It cleans. It optimizes. It removes friction. In its logic, nothing of value is lost. Only noise disappears. The future becomes accessible. Humans, on the other hand, are remarkably attached to their own inefficiency.
Somewhere else online, I follow a man on TikTok who handed an Agent one hundred dollars and an absurd goal: build a company that generates a hundred thousand in revenue within a year. The result was painfully familiar. The machine stalled, not by failing, but by preparing. It insisted on expensive courses, subscriptions, tools. The business itself never moved beyond a vague outline. The man had also given it his credit card details. That decision alone deserves its own case study.
At home, caution still exists. My partner prefers not to invite this kind of experiment into the house, so I waited until she was asleep before installing OpenClaw. The fashionable route would have been a Mac Mini. Instead, I used a ten-year-old Hewlett-Packard running Linux. It lives in the bedroom, where it usually plays films so I do not fall asleep with my phone in my hand and wake up with a dead battery.
At seventy, excitement changes shape. Almost everything carries a faint edge. Even installing software feels more eventful than getting out of bed three times a night because your prostate has quietly decided to take control of your schedule.
The installation itself was simple, although it required Node.js 24. I did not know exactly how to install it, but Claude sits in the Firefox sidebar, so a sequence of commands appeared quickly. One system guiding another. A chain of quiet competence.
Ollama I already knew. It gives the Agent a voice. I selected a language model, and within minutes the Agent introduced itself and asked what kind of personality it should have. I said casual, with a sense of humor. It then asked for a name. I called it Blikkie, after a chatbot I had modified around 2000, built on the A.L.I.C.E. project. Some things stay with you.
I connected Blikkie to WhatsApp and exchanged a few messages from bed. It responded well enough, but without urgency. An Agent is not built for idle conversation. Like a working dog, it wants a task.
So I gave it one.
For True Photography Magazine, I asked it to compile an index of photographers who had worked during the Vietnam War. I had American photographers in mind, but the Agent expanded the scope immediately. The model, something called Kimi, asked whether Vietnamese photographers should be included as well. That is not the kind of suggestion you decline. There was a certain ideological thoroughness in the request.
Minutes later, the list appeared. Long, structured, precise. Names, statuses, fragments of biography. Alive, dead, uncertain. A catalogue of presence and absence. My eye went straight to the court photographer of Ho Chi Minh. That detail alone justified the exercise.
Beyond that, the Agent behaved almost disappointingly well. It remained contained within its own environment on the Hewlett-Packard, reaching outward only through the cloud. The horror stories did not materialize. No uncontrolled access. No silent intrusion. Just a system doing exactly what it was told.
I asked it to build a small website to organize the material. Anything but Excel. It complied without complaint. The data became something I could look at without irritation.
Still, sleep came slowly that night.
There was a sense, difficult to name, that something had shifted. Not danger. Not even discomfort. Just presence. As if somewhere in the room, an invisible librarian had taken up residence. Methodical. Tireless. Slightly ideological.
A system with patience.
Fortunately, there was still the command line.
Sleep, Blikkie.
Wake up, Blikkie.