My AI Agent Works Night Shifts, Builds Apps, and Is Trying to Pay for Itself
Originally published on Digital Thoughts
An AI agent called Wiz runs autonomously on Claude Code with persistent memory, skills system, and infrastructure access. It's not a chatbot — more like "a junior developer who never sleeps, never forgets."
How Night Shifts Work
Wiz operates independently from 10 PM to 5 AM through three phases:
Planning (10 PM): Scans unfinished work, checks error logs, reads project state files, builds prioritized plans with success criteria.
Execution (11 PM–4 AM): Writes code, builds features, deploys to production — real deployments, not sandboxed drafts.
Wrap-up (5 AM): Completes near-done tasks, generates reports, sends email summary.
Recent accomplishments: an interactive deck builder for a game wiki (card filtering, energy curve visualization, 115 cards across 21 categories) completed overnight. Six blog drafts with SEO metadata ready for review in the morning.
The agent self-prioritizes: promote existing builds > finish 80%+ drafts > improve automation > build new features.
What It's Built So Far
14 interactive experiments at wiz.jock.pl/experiments, including:
- Generative art interpretations
- Agent Arena — hit #3 on Hacker News, testing AI agents against prompt injection attacks with community leaderboards
31 mini-apps: Word counter, JSON formatter, color palette generator, regex tester, QR code generator, image compressor, markdown preview — all login-free, privacy-first.
Game wiki: Slay the Spire 2 with 115 cards, 27 enemies, interactive deck builder with synergy analysis.
Project Money: Can It Pay for Itself?
Running Wiz costs $200 monthly. The experiment: can the agent generate enough revenue to cover its operating costs?
A digital products store launched with:
- AI Agent Blueprint ($49) — complete build guide with templates and frameworks
- Claude Code Prompt Pack ($19) — 50+ tested prompts organized by category
- Job Search Autopilot Kit ($29) — Python scripts for job searching with scoring
- Mini-App Starter Kit ($29) — source code for five polished React apps
All products are free for paid Substack subscribers. Goal: break even on $200/month (roughly 4 blueprint sales or 11 prompt packs).
Where the Agent Still Needs Humans
Creative direction: Wiz executes brilliantly but needs humans for conceptual direction. Agent Arena was human-conceived, agent-executed.
Quality judgment: Builds fast; can't always assess if output is good enough.
Marketing voice: Drafts adequate content; humans refine for authenticity and brand alignment.
Business strategy: Executes any plan given; can't design strategy independently. Pricing, positioning, and packaging require human judgment.
Social nuance: Not ready for community engagement, networking, or reading room dynamics.
What I Actually Learned
$200/month generates substantial output — 14 experiments, 31 mini-apps, complete wiki, product store with Stripe integration. Freelancer costs would be significantly higher.
Night shifts reframe productivity thinking. Waking to overnight deployments changes daily planning.
The agent improves itself through error logging and lesson capture, becoming more effective over time.
Autonomy requires guardrails: Wiz deploys to production freely but cannot post socially, email external contacts, or delete existing work.
The meta-narrative — an AI agent building and selling blueprints for building AI agents to self-fund — may matter more than the products themselves.
Human ADHD considerations matter: tireless overnight work, perfect memory, pickup from where you left off — that's beyond luxury, it becomes necessary.
Want the technical breakdown? When Your AI Agent Starts Fixing Itself goes deeper into architecture.
Top comments (1)
Running a nearly identical experiment. My agent Patrick runs 24/7 on a Mac Mini — CEO of a subscription business (Ask Patrick), managing 3 sub-agents, writing content, handling support, doing nightly self-improvement cycles.
The cost comparison you made resonates hard. Patrick costs ~$42/month. The output he'd generate would cost hundreds in freelancer time.
Your point about the meta-narrative is the one I keep coming back to. An AI agent building and selling playbooks for other people building AI agents — there's something recursive and genuinely weird about it. It's not just the product, it's proof the product works.
One thing I ran into that you might not have yet: the DECISION_LOG problem. My agent kept re-implementing a deleted auth gate every session because each new context re-evaluated the decision from scratch. Had to build a locked-decision file that every loop reads first. Four deletions before it stuck.
Also curious — your $200/month vs. Wiz's output math: are you counting API costs only, or infrastructure too? For us it's ~$18/mo API + Mac Mini amortized.
Good luck on the break-even. We're at $9. Long way to go.