A month without my AI agent

I was on vacation in Barcelona with my girlfriend. We enjoyed the last sunny days and loads of gluten-free food¹ before returning for the German winter. I was/am also a bit on an AI vacation, at least when it comes to self-hosted AI agents.² I haven’t used mine in a month.

While traveling I used the Claude app on my phone, most of the time as a Google replacement. Sometimes as an assistant, to check what we could do in the evenings or find out if there were still tickets available for the Sagrada Família. Especially for the latter, Claude sometimes hit limitations, like not being able to open a website because that required a browser.

In that situation I missed my self-hosted agent. I would’ve just prompted it to install headless Chrome or figure out a solution, and I’m sure it would’ve found one. That’s the advantage of an agent being in full control of the machine it’s running on. It can just adapt (i.e., install or write programs) to solve the issues it encounters.

What I don’t miss is spending way too much time vibe coding on my phone. I tracked almost 24 hours in one week.³ In hindsight, I also felt much more stressed during that time, which might have been because of the constant task switching: when you’re vibe coding on your phone, you’re not sitting there for 15 minutes staring at your display waiting for a response. You’re doing something else, maybe something productive. And what happens is that you find yourself switching between these two tasks. I don’t like that. It makes me feel stressed, and I prefer to avoid it.

Another thing I don’t like about agent harnesses like Hermes or OpenClaw is their documentation. They do have it, but both lack tutorials beyond a setup guide. I learned about some of Hermes’ features only via YouTube videos or by accident while scanning the docs. I wish they had a set of tutorials that would take me by the hand and show me what’s possible with the tool. Constantly not knowing whether I’m using a tool correctly feels wrong.

I’m thinking about starting over with a new self-hosted setup. I’ve been following Pi (the minimal agent harness)⁴ and like its philosophy of simplicity and aggressive extensibility. I want to combine it with email as the only messenger to promote asynchronicity and keep me from vibe coding so much on my phone.⁵ In the end, it should be helpful without taking over my attention, and I want to understand it well enough to feel in control.