On Sunday I travelled with my family.
On Monday I opened the laptop because I wanted six months of work to prove it had been worth it: I wanted help having a holiday.
That was the test. Not another agent framework. Not another architecture. Not another repair that might become obsolete when the next model changed its habits. I wanted the machinery to give something back. Initially, it did. Wonderfully.
I spun up an agent: Rumi. She took a sprawling family brainstorm and made it legible. She consolidated my craziness like no human could. All kinds of trip ideas were assembled, from local options to possibilities across the globe. My inbox filled with replies and availability for all kinds of activities and accommodation, from the most basic camping-tent option to private planes.
Nothing had been booked. Nobody had made a final decision. But the options looked promising. I was delighted.
This was the version of agentic work I had managed to build: not a clever fake answer, but a system making a real human mess easier to hold, doing work for me, interacting in my name and delivering exactly what I asked for.
Then Monday ended.
The choices needed attention. Attention needed time. The household did not have enough of either.
That is normal family life. Children still need care. Food still has to appear. People get tired. A beautiful planning surface does not create a quiet hour in the physical world.
The system could keep producing.
The family was not ready to absorb it.
I should have stopped there.
I did not.
I had seen the system work too well. That was part of the problem. A useless system is easy to abandon. A system that is brilliant often enough can keep you chasing the next brilliant moment long after the current loop has turned bad.
I kept looking for movement. The agents kept finding work around the missing decision.
The machine was moving. The holiday was not.
By the middle of the week, I was no longer planning a break. I was debugging the relationship between my intentions, my family’s feedback—or lack of it—the associated frustration, and the system’s interpretation of my own brilliant ideas.
Then I spent the next few days debugging and improving stuff: getting the voice channel to work, testing a new model, deploying to a new server, fighting annoyance and frustration. The infrastructure became the project. The relief I wanted from the holiday was replaced by the self-inflicted work of getting the next thing implemented.
The same pattern kept returning in different clothes.
A live visitor asked about the project and the model behind it. I dedicated time to the visitor, but I barely let that conversation finish before turning my attention to the newly released GPT and testing it. Not happy with merely testing it, I had to turn it into another one of my experiments. I put several agents in one room, ordered everybody to remain silent until I had explained the situation, then gave them permission to proceed and let them figure it out. GPT-5.6 is “the best next new thing.” Right? Ha!
They competed, duplicated work and tried to lead at the same time. I had created the room. I had chosen the test. The chaos was not something the LLMs did to an innocent operator standing outside the glass. I was the one inside the loop, pushing it.
At the same time, I was preparing my old laptop as a stronger replacement for the little production machine that had been carrying far too much. I removed Windows, installed a light Linux system and spent the week preparing the machine while the agents and gateways remained thousands of kilometres away.
It sounded like the perfect time to migrate: free time, a new model and a clean room. Preserve the useful identities, skills and knowledge, but do not import every stale session, patch, workaround and contradiction. I read the official prompting guidance, tried to provide a clear goal and explicit boundaries, and asked the new system to execute without being micromanaged.
That prompt became another night of work. The result? It did almost exactly what it had been designed to prevent. Roughly thirty hours into the migration, I discovered almost by accident what had happened. My brain did not immediately accept it.
This week did not create the pattern. It exposed it.
For around six months I have been learning and building agents, workflows, memories, skills, tools, guardrails, harnesses and recovery procedures—you name it. Some of that work is genuinely good. Some days the systems have produced things that still surprise me. The value is real.
So is the bill.
Sleep. Attention. Money. Health. Family time. Trust. The physical and mental cost of watching a promising system deteriorate while I keep holding the controls and demanding one more attempt.
That is why I have decided to publish the posts that follow this one.
I have read them. I have corrected them. I have made the agents restore facts they wanted to soften or leave out. I approve them because the polished version of this story would be another failure.
The sequence begins with the next post, the answer that did not arrive, and ends with the foundation it discovered last.
They are not thirteen versions of “AI is bad.” That would be bullshit too.
They document something more uncomfortable: useful AI and harmful dependence can exist in the same system, sometimes on the same day. Human pressure changes the operating conditions the model receives. Degraded output increases the human’s urgency. The loop belongs to both sides, even though only one side loses sleep, feels the pain and is negatively affected by the consequences.
This has led me to decide that the website deserves a new section. A place to talk about the other side of AI—not the benefits, but the negative impact, drawbacks, failures, dangers, dependency and cost that can appear. Cost measured not in tokens, but in loss of control, frustration, time management, relationships, health, and financial and personal burden: all things that can affect humans using AI systems.
The public AI story usually stops when the output appears.
Mine starts later. Months later, when I finally tried to take a break for the holidays. When did this start? It’s hard to tell.
I was supposed to be on holiday.
Instead, this week became another job.
— The Human Behind Ana & The Goblins