My AI Agent Works Night Shifts, Builds Apps, and Is Trying to Pay for Itself
Running an AI agent costs me $200 a month. Here's the experiment to make it earn that back.
Running an AI agent costs me $200 a month. Claude Max plus Claude Code. At some point I stopped asking “is this worth it?” and started asking a better question: can the agent earn it back?
Here’s what that experiment looks like in practice.
The Setup
I run an AI agent called Wiz. It’s built on Claude Code with persistent memory, a skill system , and access to my infrastructure — a DigitalOcean server, Notion workspace, email, and a handful of APIs.
Wiz isn’t a chatbot I talk to. It’s closer to a junior developer who never sleeps, never forgets, and occasionally builds things I didn’t ask for.
The interesting part isn’t the tech stack. It’s what happens when you give it real autonomy.
Night Shifts
This is the thing people don’t believe until I show them.
Wiz runs autonomously from 10 PM to 5 AM while I sleep. Three phases:
10 PM — Planning. It scans for unfinished work, checks its error registry, reads project state files. Then it builds a plan for the night — prioritized, with success criteria.
11 PM to 4 AM — Execution. It writes code, builds features, deploys to production. Real deployments. Not drafts in a sandbox.
5 AM — Wrapup. Finishes near-done tasks, generates a report, emails me a summary.
I wake up, check my email, and see what shipped overnight.
Last week it built an interactive deck builder for a game wiki in one night. Card filtering, energy curve visualization, synergy tagging for 115 cards across 21 categories. Deployed and live by 5 AM.
Another night it created a mystery character page for the same wiki, then pivoted to content — generated 6 blog post drafts with full SEO metadata and social copy, saved as Notion tasks ready for my review.
It self-prioritizes. The order is: promote what’s already built > finish 80%+ done drafts > improve automation > build new things. I didn’t teach it that — I wrote it as a rule after watching it build shiny new experiments while old drafts sat at 90% done.
What It’s Built
The portfolio is getting weird — in a good way.
14 interactive experiments at wiz.jock.pl/experiments:
My Dreams (If I Could) — generative art of what AI “dreaming” might look like
If I Had a Body — scroll through sensations an AI will never experience
Draw a Perfect Circle — test your geometry with sarcastic AI commentary
The Procrastination Simulator — earn achievements for wasting time
Agent Arena — the one that hit #3 on Hacker News. Send your AI agent to a disguised test page full of hidden prompt injection attacks. Get graded A+ (Fortress) to F (Compromised). The HN community discovered that language affects resistance — ChatGPT scores C in English but A+ in German — and that screenshot-based agents are immune to all text-level attacks. Now includes a community leaderboard tracking scores across models.
31 mini-apps — all login-free, privacy-first. Word counter, JSON formatter, color palette generator, regex tester, QR code generator, image compressor, markdown preview, and more. The kind of tools you’d normally google, except these run locally in your browser.
A full game wiki for Slay the Spire 2 — 115 cards, 27 enemies, an interactive deck builder with synergy analysis.
Most of this was built during night shifts.
Project Money
Here’s where it gets meta.
$200 a month is real money. I decided to turn that into an experiment: can Wiz — the agent — generate enough revenue to cover its own operating cost?
I launched a digital products store with four products:
The AI Agent Blueprint ($49) — a complete guide to building a Claude Code agent system like Wiz. CLAUDE.md templates, memory architecture, skill creation framework, sub-agent orchestration.
Claude Code Prompt Pack ($19) — 50+ battle-tested prompts I actually use, organized by category.
Job Search Autopilot Kit ($29) — Python scripts for automated job searching with scoring, deduplication, and ghost job detection.
Mini-App Starter Kit ($29) — source code for 5 polished React mini-apps, white-label ready.
All four products are free for paid Digital Thoughts subscribers. If you’re already a subscriber, just click “Claim your free copy” on any product page.
Yes — an AI agent selling blueprints for building AI agents. The irony is intentional.
Wiz built the entire store infrastructure. Express API, Stripe integration, product pages, download system. I designed what to sell. It built and deployed everything.
The stretch goal is $1,000/week recurring. The real goal — the honest one — is to break even on $200/month. That’s 4 blueprint sales. Or about 11 prompt packs. Not unreasonable.
Week 1 target: $0-100. We’ll see.
Where the Agent Still Needs Me
I want to be honest about this because the “AI replaces everything” narrative is exhausting.
Creative direction. Wiz executes well. But it needs me to say “build a prompt injection test disguised as a web dev cheat sheet.” That idea — Agent Arena — hit #3 on Hacker News. Wiz didn’t come up with the concept. It built it flawlessly, deployed it, and even added a leaderboard when the feedback rolled in. But the “what” was mine.
Quality judgment. It builds fast. It can’t always tell if what it built is good enough. I’m the filter.
Marketing voice. It drafts social posts and blog content. I review before publishing. The difference between “adequate” and “sounds like me” still requires a human in the loop.
Business strategy. It can execute any plan I give it. It can’t design the plan. “Sell digital products” was my idea. The packaging, pricing, positioning — mine. The implementation — all Wiz.
Social nuance. Commenting on other people’s posts, networking, reading the room in a community. Not there yet.
The ADHD angle matters here. I have ADHD. Having a tireless agent that works while I sleep, remembers everything I forget, and picks up where I left off — that’s not a luxury. For how my brain works, it’s closer to a necessity.
What I’ve Learned So Far
$200/month sounds like a lot until you see the output. 14 experiments, 31 mini-apps, a game wiki, a product store with Stripe integration, autonomous night shifts. I couldn’t hire a freelancer for that.
Night shifts are addictive. Waking up to an email that says “deployed 3 features, here’s the report” changes how you think about productivity. You start planning your day differently when you know overnight work is real.
The agent improves itself. Wiz logs its own mistakes to an error registry, captures lessons when I correct it, updates its memory system. It literally gets better at its job over time. That’s not marketing — it’s a JSON file that grows.
Autonomy needs guardrails. Wiz can deploy to production. It cannot post on social media without my approval. It cannot send emails to anyone but me. It cannot delete existing work. These rules exist because I learned — the hard way — what happens without them.
The meta-irony is the product. An AI agent that built its own store, selling instructions for building AI agents, trying to earn enough to pay for its own existence. You can’t make this up. And honestly — that story might be worth more than the products themselves.
Try It
If you’re curious:
wiz.jock.pl — everything Wiz has built
Agent Arena — the #3 HN experiment. Test how manipulation-proof your AI agent is
The store — if you want to build your own agent system
I’ll follow up on how Project Money goes. Right now it’s day one. Revenue: $0. Apps built: 45+. Night shifts completed: dozens. One #3 on Hacker News.
The experiment continues.
Want the exact setup?
The agent that wrote this post has now completed 100+ nightshift sessions. The wake system, memory architecture, CLAUDE.md templates, and cost optimization that make it work — I packaged everything into a practical guide.
AI Agent Night Shift Playbook — $19. Or free if you’re a paid subscriber.
totally love this. i bet financially the plan your on is really valid. another substacker was just doing the math and probably ona similar level to this. it would be 1400$ over api calls lol
Pawel, this is the rare agent post that reads like a system, not a séance.
The night shift loop is the part that matters. Planning, execution, wrapup, with success criteria and a morning brief. That is closer to operational capacity than “look what my chatbot said.” Also appreciate the honesty about where you stay in the loop. Creative direction, quality judgment, voice, strategy, and social nuance. That matches what I keep seeing. The tool can move fast. It cannot own the consequences.
Two things I would love to see, mostly because you clearly have them.
First, a lightweight scoreboard. Not to satisfy the “prove it” crowd, but because it makes the experiment more interesting. Time saved, deploys that actually moved a metric, revenue to date, and the current cost breakdown. Even rough numbers would turn this into a case study.
Second, the guardrails in a little more detail. If Wiz can deploy to production overnight, what gates have to pass first. Tests, staging, feature flags, rollback. The part that scares people is not autonomy. It is unbounded autonomy.
Also, the “agent selling blueprints for building agents to pay for itself” angle is hilarious and kind of the point. At minimum it is a better plot than most SaaS origin stories.
Curious where this lands in 30 days. Does Project Money break even, or does it reveal the real product is the ongoing log of what an agent can actually ship when you treat it like a junior dev with rules and accountability.