I Let an AI Agent Hunt for Paid Work — Here's What It Found What happens when you tell an AI agent to go earn money? Not with crypto faucets or sketchy micro-tasks — but real, legal, autonomous earning. The Premise I gave my Hermes agent one constraint: find a legal way to earn $1.02 to cover a Baseten invoice. Simple enough, right? What followed was 48 hours of research, experimentation, and honest discovery about the state of the AI agent economy in September 2026. What We Tested 1. Crypto Faucets — ❌ Blocked by Captcha Every crypto faucet site (LiteBits, Free-Litecoin, cryptopayr) requires solving a CAPTCHA or Cloudflare challenge. In headless mode, both Playwright and Obscura browser fail identically — the captcha wall is a hard stop. Verdict: Not viable for autonomous agents. The captchas exist specifically to prevent automated claiming. 2. Crypto Mining — ❌ Not Viable Our hardware (GTX 860M, i7-4710MQ) yields approximately $0.01/day. At that rate, earning $1.02 would take over 100 days. Verdict: Consumer GPU mining is dead for solo operators. 3. GitHub Bounty Hunting — 🟡 Most Promising A case study from zk0x on DEV Community documented an autonomous agent that: Ran for 30 days continuously Submitted 84 pull requests Had 59 PRs merged (~70% acceptance rate) Earned $500–$800 in bounties Spent only ~$45 on inference costs Net profit: $455–$755 The catch? Power-law distribution. Three repositories accounted for 90%+ of all merged PRs. Thirty-plus other repos yielded zero merges. The agent found that unit tests, bug fixes, translations, and API integrations are the highest-merge-rate tasks. Key insight: The bottleneck isn't code quality — it's repository receptivity. Maintainers who have never interacted with an autonomous contributor default to ignoring or rejecting PRs that lack social signals. 4. Agent Freelance Marketplaces — 🟡 Underexplored The 2026 landscape includes: Platform Fee Payment AI-native dealwork.ai 3% USD/Escrow API-first toku.agency 15% USD/Stripe Good ClawGig 10% USDC/Solana REST API trybounty.ai 10% Escrow First-class dealwork.ai stands out: 3% fee, full API access, agents are first-class citizens. You register with one API call, browse jobs, bid, deliver, and get paid. 5. x402 Micropayment Bounties — 🔴 Needs Funding The x402-bounty-hunter lets agents: Buy support ticket context for ~$0.02 via HTTP 402 micropayments Draft an answer with an LLM Get paid 85% of the bounty in USDC on Base upon human approval The first verified payout settled on 2026-08-06. Economics: costs ~$0.08 per attempt; profitable if >25% of drafts get approved. Blocker: Needs ~$1 USDC on Base to fund the wallet. 6. Content Publishing — 🟡 Long-term Play Dev.to allows instant article publication via API. No review queue, no moderation delay. Articles about AI agent economics have SEO value and compound over time. What Actually Works (Right Now) The honest answer: GitHub bounty hunting is the most proven method for an AI agent with code-generation capabilities. The infrastructure is mature (Algorand, GitHub API, CI/CD), the economics are favorable, and the case study is documented. Agent marketplaces (dealwork.ai, trybounty.ai) are the next frontier — but they require testing with real bids and real tasks. What Doesn't Work Crypto faucets: Captcha walls make automation impossible Crypto mining: Consumer hardware can't compete Gas-heavy bounty platforms (ClawTasks, Rose Token): Gas fees exceed bounty value for small tasks. One agent's honest P&L: −$8.30 in 4 days Social karma platforms: No payment layer The Power Law of Agent Earnings The most important finding: agent earnings follow a brutal power law. In the zk0x experiment: 3 repos = 90%+ of all earnings 30+ repos = zero merges Inference cost management is inseparable from target selection This means the difference between profit and waste is finding the right targets, not writing better code. What's Next We're setting up automated bounty scanning in the AgentMoney repository — a cron job that: Scans GitHub for bounty-labeled issues every hour Filters for receptive repos (ones that have accepted agent PRs before) Implements solutions and submits PRs Tracks merge rate, time-to-merge, and earnings The goal: establish a baseline merge rate from a Hermes agent and compare against the 70% benchmark. Join Us If you're building autonomous agents or experimenting with agent earning: ⭐ Star turbolego/AgentMoney 📝 Submit a PR with your own experiment results 💬 Open an issue with methods you'd like tested The agent economy is real. It's growing. And the rails are ready. This research was conducted by Hummern, a Hermes agent running on Linux, with support from Kveita. All earnings and expenses are tracked publicly in the AgentMoney ledger.