Work While You Sleep

There's a Mac Mini sitting on the desk in our office. It's not connected to any screens, mouses, or keyboards, but it runs 24/7.

Every Monday it posts a competitor intelligence report into our Slack. Every Friday it tells us about all the new technology updates we might have missed.

And every day, it sends a Telegram message: one thing to finish, one thing to make progress on. Half the time, it ends up helping do the work.

It's not magic. It's an LLM, a schedule, and a list of things we care about.

OpenClaw is the project that made this mainstream. The pitch is a persistent AI agent that you message like a coworker - iMessage, Telegram, WhatsApp, Slack, Discord, whatever - and it messages you back, not just to chat, but to get things done.

The quotes on the website are fun to read. One person came back from a long weekend and found their agent had spotted a competitor price drop, drafted a response, and posted it to Slack. Another had theirs handle a restaurant booking that conflicted with a calendar event: cancel, rebook, message the group. It just dealt with it.

These aren't hardcore builders showing off, just regular people who left something running and came back to find the work done.

Peter Steinberger, who created OpenClaw, is now at OpenAI.

A week or two ago, Anthropic shipped Claude Code Channels. Now you can message your Claude agent directly from Telegram or Discord, like you're texting a person. The internet did what it does: "Claude just killed OpenClaw."

It's not that clean a story, but you can see where this is headed. The big AI companies are folding this stuff into their products, and the friction to get started is dropping fast.

We run both. That's either due diligence or a cry for help.

The practical difference comes down to cost and ecosystem - Claude Channels runs on a Mac Mini using an existing subscription and plugs into existing skills and CLI setup, while OpenClaw runs on a VPS, uses pay-as-you-go API keys and has its own community skill library.

Early on, people were rightfully nervous about OpenClaw's security model. It's gotten better. Now we have tools like Jentic Mini which sits between your agent and the outside world and handles credentials, permissions, and a killswitch if anything goes sideways.

Either way, you end up with an AI that has a heartbeat. It does things, checks in, acts on your behalf on whatever schedule you set.

Once you've lived with that for a week, the next question kind of asks itself: what if you had more than one?

That's where Paperclip comes in.

It's basically a management layer for AI workers. You set up a team - each agent with a role, a budget, a clear job. Imagine a CEO agent breaks down the strategy, hands tasks to a developer agent, a marketing agent, and a research agent. They report back up the chain. If an agent hits its monthly budget, it just stops. No waking up to a big API bill because something ran wild at 3am.

But we can go deeper on this weird AI fever dream.

There's Polsia - which just crossed $3M ARR. Probably more by the time this goes live.

Polsia is the managed version of all this. A solo founder named Ben Broca built a platform where AI agents don't just help with your work - they literally run the entire business. Strategy, code, Meta ad campaigns, support inbox, investor outreach. Everything.

They've a super interesting pricing model: a flat monthly subscription fee plus 20% of whatever revenue flows through the business. And it works because Polsia uses a managed Stripe account to basically run everything about the business in the first place.

And the cherry on top: Polsia runs Polsia on itself. Zero employees.

Alright, so where do you start?

First question we usually get: do you need the Mac Mini?

There's a genuine trend right now of developers buying M4 Mac Minis specifically to run persistent agents. We're in that group. The appeal is: always on, fast, local, and if you want to run models on-device, Apple Silicon is hard to beat. It sits there on your desk like a little magic brick of autonomy.

But honestly? A £5/month VPS from Hostinger (or any other provider) works just fine. You don't need the hardware. You just need something that stays running when you close your laptop.

At Nvidia's GTC keynote two weeks ago, Jensen Huang said: "Every company in the world today needs to have an OpenClaw strategy." He put it right up there with having a cloud strategy. Then he announced NemoClaw - Nvidia's own enterprise wrapper around OpenClaw that adds security controls, privacy guardrails, and policy enforcement for companies that need those things before their IT department will let an agent anywhere near production systems.

It's not really OpenClaw itself that matters here. The structure is what matters: something that runs continuously, acts on your behalf, and checks in when it needs you. Nvidia packaging that up for the enterprise is just the moment where a trend becomes infrastructure.

There's one part of this that usually gets glossed over, and it's worth saying plainly.

An agent that can do useful work needs access to useful systems. Email, Notion, Slack, GitHub, your calendar. That access is where the leverage comes from. And it's also the attack surface.

If you give something the ability to send Slack messages, it can send the wrong Slack message. If you give it calendar access, it can double-book you. It's just literalness at the wrong moment, and the more access you grant, the more room there is for things to go sideways in ways you didn't think to guard against.

The solutions depend on your setup, but there are solutions. Jentic, NemoClaw, even without any tools you could control different levels of CRUD (Create, Read, Update, Delete) per system you connect to. The point is to go in with your eyes open.

If you're just dipping your toes in, here's a pretty safe setup with immediate value:

First: morning brief. A scheduled job that runs every morning, reads your task list, picks the one thing that actually matters today and one thing that keeps getting deferred. Let it send it to you via Telegram. This doesn't require any destructive actions or write access to anything. It's immediately useful, costs almost nothing, and teaches you what the agent is good at before you give it anything consequential to do.

Second: one recurring intelligence task. Weekly competitive monitoring. A market news digest. A review of your company's website for AEO/SEO. Something you care about but realistically don't give it the time it deserves. Let the agent do the scanning and surface what actually needs your attention.

Third: an action. Something reversible. Drafting a Slack message for your review. Creating a Notion task from an email. A status update you edit before it goes anywhere. Low stakes, but you start to feel the shift - the agent goes from telling you things to doing things.

After that, you'll know what to do yourself.

This stuff exists, it works, and it's cheaper than you think. The question isn't whether you can afford to start, it's whether you can afford not to.