Rebuilding this site from the middle of the Irish Sea
An afternoon on the ferry, an iPhone, Starlink, and an AI agent doing all the typing.
This morning I sat in the ferry lounge heading over to watch Liverpool, nursing a coffee and my iPhone. Somewhere in Frankfurt, a VPS I hardened last weekend was waiting for a task. Between me and that server was a Telegram thread. By the time I disembarked, I'd done something that would have taken a laptop and a proper workspace three times as long. This is what that looked like.
What changed
Two weeks before this trip, I rebuilt the VPS from the ground up. Not because anything was broken. Because I wanted to know what "hardened" actually means when you're running agents that touch your infrastructure.
Ubuntu 24.04 clean install. SSH key auth only, port 22 is the only open port. Unattended-upgrades so security patches don't pile up while I'm not looking. Fail2ban to bury brute-force attempts. A second user account with sudo for the agent work, keeping root for actual admin. An env file with the Anthropic API key, mode 600, so it's not readable by anyone else on the box. It's not hardened by serious standards. It is hardened by "person with a phone" standards, which is the bar that matters here.
The agent stack itself is OpenClaw running Claude Haiku, parked in Frankfurt on a Hetzner box that costs about what a gym membership does. The agent has SSH access to that hardened box. It can read files, write files, check running processes, restart services. It cannot sudo. It cannot access the API key directly. It can only do what the skill file explicitly allows.
How it actually worked
The workflow was unglamorous. Me sitting on a bench with a ferry-strength coffee. Telegram on my phone. A series of short messages describing what I wanted. A few screenshots of the existing site I wasn't happy with. The agent digesting that, asking clarifying questions through the same thread, proposing directions.
When I said yes, it didn't ask me to wait. It issued commands over SSH. Built the file structure. Wrote the HTML. Pushed it live. Reported back what it had done and where to look if I wanted to tweak anything. I looked. Sent back notes. It made the edits, cleared the Cloudflare cache so the new version appeared instantly, and reported back with the URL.
This would be mundane if I had a laptop. On a phone it's the difference between feasible and impossible. You can't code on an iPhone for hours. You absolutely can describe what you want, react to drafts, and issue corrections in a thread. That's the actual shape of how this works.
The key part nobody talks about is constraints. Because the agent is constrained in what it can do (only the commands in the skill file), I can be confident it won't accidentally delete the wrong directory or SSH into the wrong server. Because it's running on my infrastructure, not a SaaS platform, there's no rate limit and no usage cap. Because Telegram is just a chat interface, there's no friction between thinking of something I want changed and telling the agent to do it.
What this taught me
Mobile work is viable for serious technical tasks if you've constrained the system properly. I didn't spend the ferry ride debugging. I spent it describing, reacting, and making judgment calls. The work that actually needs a human (deciding what something should do, evaluating drafts, catching things that don't match the brief) sits naturally in text exchange. The work that needs a machine (writing the code, running the commands, managing the files) I handed over.
The VPS hardening mattered more than I expected. Because I'd already thought through what the agent could and couldn't do, I didn't hesitate to give it real SSH access. I wasn't nervous about it. It couldn't run arbitrary commands. It couldn't read my API key. It could only do what I'd explicitly allowed in the skill file. That's not paranoia. That's just the price of automation you can actually rely on.
Telegram is underrated as an operational interface. There's no app to open on your device, no dashboard to load, no learning curve. Just messages. I've seen people build elaborate chatbot dashboards that end up being slower to use than a text channel. This is faster. When something breaks at 3am and I'm half asleep, I want to tell a thread what to check, not navigate a UI.
The constraint of a phone forced useful discipline. You don't write verbose instructions on a tiny keyboard. You say what you want. You wait. You react to what comes back. It's shorter feedback loops than you'd have with a laptop because you're too uncomfortable to overthink it.
Where this actually sits
This only works if you're willing to give the agent real permissions and trust the constraints you've built. If you're nervous about letting an automated system touch your infrastructure, you shouldn't. But the nervousness usually isn't about the system. It's about not having thought through the constraints clearly enough. Spend an afternoon on VPS hardening and permission scoping and the nervousness goes away because the risk is actually low.
It only works if you know what you want. I can't describe a design to an agent on a phone if I don't already have a point of view. The agent is good at execution, excellent at questions, terrible at invention from nothing. You bring the judgment. The agent brings the speed.
You lose certain things working from a phone. You can't keep six browser tabs open while you're thinking. You can't do the kind of flow state work where you're writing code and debugging in real time. For that you need a proper keyboard and a real screen. But for everything else, for architecture and decision-making and reaction cycles, you lose almost nothing. And you gain the ability to do serious work from places serious work wasn't supposed to happen. Like the lounge of a Liverpool-bound ferry, where the agent and I had a far more productive morning than the team did an afternoon.
The actual takeaway
I got off the ferry with a new site live, this post drafted, and the distinct feeling that I'd solved a problem I didn't know I had. The assumption in tech is that serious work requires serious equipment. A laptop. A desktop. An office. But if you've hardened the infrastructure and constrained the agent properly, you can do serious work from a ferry bench with a phone and a thread.
This isn't about AI being magical. It's about the constraints working. The agent can't do what I don't explicitly allow. I can't overthink things on a phone keyboard. Telegram is just a chat interface, so there's no friction getting from thought to action. The VPS is my infrastructure, so there's no SaaS platform getting in the way or rate-limiting how much I can ask for.
It's a workshop that travels. And unlike every other time you've heard someone say that about their tools, this one actually delivers.