Is my AI agent useful — or just a Tamagotchi eating my time?

My AI agent was supposed to take work off my hands. That was the plan, anyway.

I'm a self-employed video producer. I develop ideas, shoot, edit and talk to clients. At the same time, I deal with everything the romantic idea of creative independence tends to leave out. Appointments. Emails. Files. Quotes. Project updates. New business. Another file. And somewhere there is always a piece of information I know I saved. I just no longer know where.

When the first usable AI agent appeared with OpenClaw, I immediately pictured a small computer taking over the organisational side. A personal assistant who knew my projects, connected the information and gave me the right thing at the right moment. I would look after the films. She would stop my business disappearing between emails, project folders and half-finished to-do lists.

So I set out to build that assistant myself. Things escalated quickly.

I'm not a programmer. I'm not a computer scientist either. Before this experiment, I rarely had a reason to open a terminal. Now I was supposed to type commands into one, manage Linux, configure access and let an AI system work on a computer.

A sensible person would probably have started with one small task.

I started by giving my agent a name and a job. She's called Trix, and I wanted her to be more than an assistant. In my head, she was the organisational lead of a virtual production company. She would know every project, create files, bring together knowledge about different clients and watch the whole engine room while I made the creative decisions out front.

She can read documents, access folders and carry out tasks. That was exactly what I'd been missing from earlier AI chats. I didn't want to copy old conversations into new windows forever. Trix should know what we're working on.

An agent with broad access to your computer is useful. It's also a very good reason to stop and think about security.

A system that can operate the terminal can do a great deal on that machine. If you don't fully understand what's happening there, you develop an interesting mixture of megalomania and mild panic.

Luckily, I still had a ten-year-old Windows laptop. I wiped it, installed Linux and set up OpenClaw there. Then I connected Trix to anything that sounded useful: Discord, Telegram, Gmail, Calendar, Google Docs, Sheets and Drive. External tools followed. Separate access seemed safer than handing her a clear path into my digital life, so Trix got her own Google account, email address, calendar and Drive. I forward relevant work emails to her. When she creates an appointment, I see her calendar inside mine. My private email and calendar stay outside.

That separation was one of the first decisions that later proved genuinely sensible.

First Trix needed an identity. Then a memory. Then abilities — skills. Then separate sessions and access to my files. So far, this was still close to what I'd originally wanted.

I chose Notion as the interface. The problem was that I'd hardly used Notion and didn't particularly want to. Suddenly I had to learn a new tool so my agent could make my existing tools easier.

I made a channel where I sent Trix articles and information. Read this. Learn that. Install this skill. Look at that system. Every discovery was supposed to make her stronger. In my head, Trix became ‘Mega-Trix’: one agent that would eventually do everything.

The AI replied very reliably: I will.

That's a dangerous answer when you have many ideas and very little knowledge of software architecture. I imagined an agent would simply begin, work day and night, then return a day or two later with the finished result. In practice, she would work for a long time, produce files, build interfaces, get tangled up and sometimes stop responding by the next morning.

My first version failed under too many ideas. For the next one, I wanted more structure and lower costs, because tokens were disappearing alongside my time. On the first day with a particularly powerful model, about €50 of my €100 credit vanished. I had apparently used a great deal of computing power to build a system that might one day save me work.

In some conversations at the time, ‘token maxing’ almost sounded like a sport: whoever used the most tokens must be especially productive. I can confirm that it's possible to use a lot of tokens. The relationship with productivity is rather looser.

Constantly switching models didn't make the system more stable. This one today, another tomorrow, a separate choice for every task. I chased every shiny new option and wondered why the structure kept falling over.

For days — sometimes weeks — I barely did the work I'd wanted the agent for. Instead of making films, I sat in front of my digital pet. I fed it skills, checked whether it still responded, rewrote rules, repaired its memory and reset it when it got stuck again.

My agent hadn't become an employee. She was a Tamagotchi.

Except the old Tamagotchi beeped when it was hungry. Trix could generate API costs while reading my calendar.

That's the uncomfortable side of agent hype. It's easy to feel you now have to build your own digital universe. Automate everything. Test another model. Add another dashboard. Send another subagent.

Building feels productive. You're always busy. You fix a bug and improve a structure. Once everything is finished, you think, I'll finally have time for my real work.

For a while, that future didn't arrive.

I wanted AI to extend my abilities. Instead, I mainly extended my to-do list with Linux, accounts, permissions, model choices, token costs and software architecture — precisely the work I'd never wanted as a career.

So I stopped judging my agent by its theoretical potential. Since then, I've asked a much more ordinary question: does Trix help me with a real project today?

The current version of Trix is much leaner. The old laptop has retired. The agent now runs on a VPS, a virtual server. Fewer connections. Fewer interfaces.

What I need now is closer to a small second brain. Trix doesn't have to do everything. She needs to be reachable, know our shared work information, and help me find or create it.

She still has her own Google account. She creates documents in her Drive and shares them with me. She adds appointments to her calendar, which appears in mine. I forward work emails instead of letting her search my entire inbox. Our shared workspace remains understandable and reasonably separate from my private digital clutter.

We talk through Discord. Not because Discord is the perfect AI-agent interface, but because I already had it and know how it works. That has become a fairly important test for me: no more new tools. Trix has to be where I am.

In the morning, I get an email telling me what's coming up. I don't need a giant dashboard to tell me I have an appointment.

For now, Trix mainly does secretarial work. My bigger ideas haven't disappeared: the virtual production company, automated video production and everything else are still there. Occasionally I show the plans to a new model and ask whether it could build them now. The answer is usually enthusiastic. Better ask again later.

I still find AI agents exciting for my work. Otherwise, I'd have stopped after version two. I just no longer think their value depends on how many systems they control or how impressive their architecture looks in a diagram.

My conclusion for now: an agent is useful when she supports the work I'm actually doing. When she finds information and prepares documents so I can get on with what I needed to do anyway.