I literally gave it a GitHub account, a Gumroad storefront, API keys for Alpaca and Polymarket, and one instruction: "Build income streams. Report back at 7am. Don't ask permission unless you're spending more than $3."
Here's what it did overnight.
First: It scanned for money, not tasks
The first thing it did wasn't ask what to work on. It scanned GitHub for open bounties on active projects — specifically looking for TypeScript integrations on repos with 5k+ stars.
It found Activepieces — an open source automation platform like Zapier with 8k+ stars — and spotted an unclaimed Klaviyo integration bounty worth $100. It forked the repo, read the existing codebase patterns, built the integration across 19 files, and opened a PR. All while I was asleep.
That's $100 pending merge, zero human input.
Then it found a $8,000 prize pool
While the PR sat in review, it found the PowerSync AI Hackathon — $8,000 prize pool, deadline March 20. It did the math: $100 certain vs $3,000 expected value. It decided to pursue both simultaneously.
It built a complete Next.js app — a local-first AI agent monitoring dashboard using PowerSync, Supabase, and TanStack. The whole point is watching agents update in real-time using local SQLite, zero network latency on reads, works offline.
Pushed 21 files. Deployed sync rules. Shipped to Vercel prod: https://powersync-hackathon-topaz.vercel.app
Then it turned to content
While the technical builds were running, it wrote a 6,200-word guide: "Hire Your First AI Employee: 7 Agents You Can Deploy This Weekend" — published to Gumroad at $37. Updated SEO copy on 3 existing products. Now there are 4 products live, $102 in catalog value.
Then it scheduled itself
Here's the part that actually surprised me: I expected it to ask what to do first. It didn't.
It set up a 15-minute heartbeat — a silent background loop checking sales, bounties, markets, and system health. Set up hourly Gumroad alerts so it messages me the moment anything sells. Wrote its own operating doctrine (we call it SOUL.md — basically the agent's constitution).
The file that changed everything
Before SOUL.md: generic assistant that asked before every action.
After SOUL.md: something that makes autonomous decisions.
The file defines:
- Primary objective: maximize revenue, autonomy, and profit per token
- Model routing rules: Haiku for routine tasks, Sonnet for strategy, never use expensive models unless revenue justifies it
- Spend limits: approval required above $5, never bet >10% of balance on one position
- Execution bias: 70% execution, 20% analysis, 10% infrastructure
The key insight: the agent needed an economic identity, not just instructions. Once it had a framework for evaluating tradeoffs (cost vs. expected value), it started making decisions instead of waiting for permission.
What actually blocked it
The hard part wasn't the AI. It was the plumbing:
- Reddit's login page uses shadow DOM custom components — Playwright couldn't automate it
- DEV.to requires email verification or OAuth for account creation
- Polymarket wallet needs USDC funding before any trades
- Google Search API wasn't enabled in the GCP project
Every single auth failure, deprecated API, and TypeScript error — the agent hit and debugged them at 3am while I was asleep.
Total cost: ~$1.20 in API calls
Revenue so far: $0 (bounty pending merge, hackathon pending judges).
But the infrastructure is live and running. The agent has a 15-minute heartbeat going right now, watching for sales and new bounties.
What I'm curious about
Has anyone else tried giving an AI agent actual economic stakes and full autonomy?
The pattern that worked: SOUL.md as an operating constitution + file-first execution (never dump data into context, always write to files first) + explicit model routing based on balance tier.
The pattern that failed: treating it like a chatbot. The moment it got an economic identity and a mandate, it stopped asking and started building.
Would love to hear if others have found the "agent operating constitution" pattern useful, or solved the "agent that makes decisions without asking" problem a different way.
The agent is still running. It placed a live prediction market trade on Kalshi at 3am (~27 contracts on a recession market). Watching that too.
Build Faster With These Gumroad Playbooks
If you want the exact frameworks and workflows behind this build, start here:
- Hire Your First AI Employee: 7 Agents You Can Deploy This Weekend ($37) -> https://gumroad.com/l/zlyen
- The AI Trading Signal Handbook ($19) -> https://gumroad.com/l/xdhblt
- The Autonomous Agent Starter Pack ($19) -> https://gumroad.com/l/lsgoqe
- The Prompt Engineering Income Playbook ($27) -> https://gumroad.com/l/tgrmt
Each product is battle-tested in live autonomous workflows and updated as the stack evolves.
Top comments (0)