Chapter I: Reflection on hiring my first 24/7 AI worker

huangpf
Feb 9
4 min read

(and why the chime on my phone is my new coworker)

A few days after John's talk at Stage: "One Person, One AI, One Product" (read it here), I had an uncomfortable realisation:

I've had a Claude Max subscription for a while, and I've never once hit the rate limit. I was using it like a tab I'd open only when I remembered.

John's sharing planted a simple idea in my head: if one person can ship production-grade software without a team, the bottleneck isn't engineering capacity. It's how you set up your workflow, your standards, your feedback loops.

the chime

So I tried a week-long experiment. I stopped treating AI like an occasional assistant and started treating it like my 24/7 employee.

That experiment is now 2 notification chimes on my phone:

Slack - Wow.mp3 → plays when Claude finishes a main task
Crystal.mp3 → plays when one of the sub-agents is done

Recently my phone has been chiming non-stop, and whenever someone asked me why it did that, I'd say, "that's my employee reporting progress."

Because once you have a sound that reliably means build completed, your relationship with time changes. You stop hovering over the terminal. You stop switching tabs to check if the code compiled. You hand off the task on your laptop, go to a cafe, and resumes on your Samsung Fold.

the setup

The setup itself is quite simple:

Hetzner cloud server for about $14 a month - On 24/7.
A private network VPN so my phone, laptop, and server all talk like they're on the same Wi-Fi.
Terminal emulator on my android so I can reach the server from anywhere.
Text-to-speech so I can listen to the responses from my Claude Code without even looking at my phone
Speech to text so I don't even have to type.

I can be driving and thinking through how I want a feature to work. I can be walking and telling the worker to scaffold the next component, refactor a module, or write tests against a spec I described out loud. The response gets read back to me. I review, give feedback, course-correct, all without touching a keyboard.

P.S. Thank you John, for entertaining my constant badgering

the night it broke & i couldn't sleep

This worker also doesn't care about office hours. If it finishes debugging at 0300, it tells me at 0300. No shame. No concept of "after hours."

There was one night the Hetzner server went down because the RAM overloaded. The worker kept trying to ping me, again and again, until I woke up and fixed it. Annoying. Also revealing. It kept going when I wasn't around and demanded maintenance when its environment failed. When I looked at the terminal, it also showed me that it had just done 80 iterations of code to hit the 25/25 score that I tasked it to do at the start of the day (More on this in the next chapters).

The part that surprised me most wasn't the convenience, but the psychological shift. I'm used to work being synchronous: output only happens when I'm seated at the desk, grinding. This system made work asynchronous. I send a task off, do something else, and let the agent run.

the economics

Some people still balk at a $20 subscription for AI tools. I am oversubscribed on AI tools.

The server is $14 a month. On top of that, I spend > $400 a month total on AI subscriptions and tools. It's not cheap (when you anchor it against a SaaS subscriptions, or netflix), but when you benchmark it against hiring, it's a steal.

For a few hundred dollars, you get someone that designs, reviews, builds, iterates, and keeps moving while you sleep. Is it flawless? No. It can be lazy. It can tick the checklist instead of solving the actual problem. It needs standards, judgement, and feedback, or it drifts towards the easiest version of done. It needs another agent to keep it in balance.

But the leverage is real. This is the first time I've felt I can buy meaningful execution capacity without building an entire organisation around it first. It really changes what one person can produce.

hanging around the community changed me

It also made me think about why SQ Collective matters right now. A coworking community sounds like a nice-to-have until you're deep in a build and you realise you're absorbing people's defaults.

Being around builders really shifted my own mental model. I watch people build and ship, take direct feedback, and keep going, and now I want to do the same. John's talk rewired an assumption for me: I don't need a team to start. I can start, ship something real, and let reality correct me.

Combine that with a 24/7 worker and the question stops being "do I have time?" It becomes "what do I build first?"

so... what's Michael building to justify the $XXX/month

You might be asking, how do you justify that $XXX/month?

The first digital employee I'm building is a digital community manager. His name is Kai.

Here's the idea - Consent-based intros, with context: talk to Kai for 10 minutes and get mutually agreed introductions to people inside the SQ Collective community, with a reason for why the match makes sense. (yes, inspired by boardy.ai for those who know).

I'm tuning the conversation flow right now, so stay tuned!

what I’m writing next in this series

Over the next few posts, I’m going to share some build notes behind Kai:

Kai’s JD: How Kai fits in within SQ Collective, and what his job is
How I am tinkering with multi-model and agent orchestration: As a non-technical folk
How I set up my AI to work while I sleep: 3 million tokens and 80 iterations to a quality bar
How I set think about designing the UX of a conversation agent: small talk, progressive disclosure, and “sounding human”
Reflections on the business model of Kai: ???

If you're building with AI, or curious about building, come hang out at SQ Collective. We run Co-working fridays and the Stage. The fastest way to level up isn't another tool. It's being around people who are shipping.

P.S. And if you want to volunteer for a 10-minute Kai test call, message me "KAI." You get introductions only if both sides agree.