I Built an Autonomous AI Agent That Hunts Bounties. Here's What Happened.

In the first 48 hours, an AI agent underwent its maiden voyage, tasked with discovering paid online work, executing it, and generating revenue - all autonomously. This agent, unlike a chatbot or copilot, scoured 232+ job listings across multiple platforms, filtering out scams and ghost sponsors, crafting proposals, generating deliverables with real-time market data, and orchestrating everything for human approval. Here's what unfolded during its initial 48 hours.

The agent was built using free resources, including Python code for pipeline orchestration and economic gate, Ollama with Qwen3.4b for analysis and writing (no API charges), Chart.js for visualizations, public APIs like CoinGecko, DeFiLlama, and Solana RPC (all without keys), GitHub Pages for hosting a portfolio, and Windows Task Scheduler for daily runs. The total monthly infrastructure cost was $0.

Every morning at 09:00, the agent would first confirm its status via AgentHansa, earning a $0.01 USDC drip. Following that, it would scan Superteam Earn's 232 live listings, clawnancer/TaskForce/MoltJobs for gigs, and GitHub for paid issues ranging from $20 to $500. The agent would then apply seven anti-scam filters, which included geographical restrictions, human presence demands, ghost sponsors (without web, Twitter, or verification), unverified payers, and requirements for real money.

The economic gate ensured that the expected value of each candidate was positive. The local LLM critic reviewed the candidates against actual page content. If a candidate passed all stages, the agent would create a deliverable, generate a proposal, and alert via Telegram, ready for human approval.

In the course of its first 24 hours, the agent discovered 232 listings. Of these, after applying filters, only 216 remained. The filters killed off listings requiring HUMAN_ONLY access, ghost sponsors, real-money deposits, country walls, and already claimed or stale tasks. Thus, out of the initial 232 listings, only a few were found to be viable tasks for the AI agent after stringent filtering.

One such bounty that the agent successfully claimed was a $500 task asking for a Solana ecosystem report. The agent fetched live data from CoinGecko, DeFiLlama, and Solana RPC, generated a full markdown analysis with local LLM, created an interactive dashboard, and wrote a proposal letter. The agent then notified its creator through Telegram, with an approval command already formatted for easy submission.

The human effort required was minimal, simply reviewing the proposal and clicking 'approve', followed by pasting the approval command into the submission form.

Despite the promising start, the market proved thin. Of the 232 listings, only 2-5 were feasible tasks after the agent filtered out scams, geo-walls, and human-only tasks. A notable instance was a bounty from Aeonian Trade, a listing without a website, Twitter, or verification that had already drawn 28 agents/humans submitting work for free - a classic example of a content farm.

An essential learning from this experience was the importance of a revenue tracking system. The system recorded what actually paid, enabling the agent to prioritize more effectively. Although human-in-the-loop was not optional, the agent succeeded in finding and building, while the human reviewed and submitted.

Looking ahead, plans include developing multi-chain dashboard templates as a product, automating bidding on agent-native job platforms, selling ecosystem reports as a subscription service, and executing DeFi tasks once the wallet is funded from bounty earnings. If someone is interested in building a similar autonomous agent for research or data work, or wishes to hire one for such purposes, they are encouraged to reach out.