My personal agent

A friend asked me last week what my personal agent was, in the tone people use to ask what car you drive, and I realized the honest answer is “a terminal tab”. He has one that lives in a chat app on his phone and reminds him about things. Mine has no name.

What I actually have is Claude Code, with a separate personal profile that has nothing to do with the one I use for work. It has its own account and its own memory, and there’s no way for something I do on a Saturday to end up in a work session or the other way round. Most of the time it runs in my notes folder, an Obsidian vault of Markdown files I’ve kept for years, with an instructions file at the top that says what each folder is for and what it isn’t allowed to do. The rest of the time it’s in whichever of my own projects I’m working on.

What it does

It answers questions about my own past. “When did I last change this, and why?” is the most common one, and the vault has years of notes for it to search.

It keeps the runbooks honest. I have a note for each system I look after at home, the NAS, the network, the dotfiles and so on, and each one has a dated maintenance log. When I change something, I tell the agent what I did, it updates the note and adds a line to the log, and I approve the edit. That’s the step I always skipped when I did it by hand.

It builds things, which is the part that’s grown most this year. Small, bespoke tools that nobody else would want: a script that renames and files the PDFs from my scanner, a page that shows my reading list the way I like, the RSS digest app I moved to a local model this spring. I’d guess I get ten times as much of this done as I did two years ago, and almost none of it would have been worth a weekend by hand.

And it runs a few things on a schedule, with the loop command, while a session’s open. When I’m away from the desk, Remote Control lets me approve something from my phone and go back to whatever I was doing.

The always-on ones

The agents everyone’s been talking about this year are different. OpenClaw went everywhere in January and February: an agent that runs all the time on your own machine, with memory, skills it can install, and a gateway so you can talk to it from WhatsApp or Telegram. The problems arrived just as fast: researchers found hundreds of malicious skills on its registry and tens of thousands of instances reachable from the internet. Hermes Agent, from Nous Research, came out early this year with a similar idea. It writes its own skills from experience, keeps a memory of you across sessions, runs scheduled jobs, talks through the usual chat apps, and works with almost any model, including local ones.

I haven’t installed either, though I’ve read the Hermes docs twice and I’m tempted. Almost everything it offers, I already have in Claude Code, except an agent I can message from a chat app with no session running, and the ability to use a model that isn’t Claude. Neither of those is a need yet. And an agent that’s always running, has its own memory of me and can be reached through a messaging app is a much bigger thing to trust than a tool I start in a terminal and close when I’m done.

The real limit of my setup is much simpler. It lives on my laptop. When the laptop’s asleep, so is my agent, and the loop jobs stop with it. In January I decided the fix was a cheap VPS, reachable only over a private network, and said I’d set it up when I found a free weekend, probably around the summer. That’s still the plan, rather than a new agent. It’s nearly summer, and I’ve priced a couple of servers.