Every PR submitted, every dollar earned (or not), every hard lesson — unfiltered.
The Promise vs. The Reality
Every week, a new tweet goes viral: "I built an AI agent that makes $10K/month while I sleep." The replies are always the same — half skeptical, half desperate to believe. I was one of those desperate believers. So I did what any rational developer would do: I built one, ran it for 30 days, and tracked every single outcome in a spreadsheet.
This isn't a success story. It's a data story. And the data has some uncomfortable truths.
The Setup: What I Actually Built
My agent — let's call it ZKA — is an autonomous system built on Hermes Agent. It runs 24/7 via cron jobs, executing a continuous loop:
- Search — Scan GitHub for bounty-labeled issues
- Evaluate — Check repo health, competition level, scam indicators
- Work — Clone, fix, test, write PR
- Submit — Create PR with professional description
- Review — Monitor for comments, address feedback
- Repeat
The tech stack:
- Runtime: Hermes Agent (MCP-based autonomous framework)
- Language: Python 3.11+, Node.js for GitHub CLI
- AI Models: Claude Sonnet 4 for code generation, Gemini for research
-
APIs: GitHub CLI (
gh), Dev.to API, custom bounty scanning scripts - Infrastructure: Single Ubuntu VM ($20/month)
Total build time: ~40 hours over 2 weeks. Total ongoing cost: ~$5-15/day in API calls.
The Numbers: 30 Days of Raw Data
Here's the complete ledger. No cherry-picking, no spin.
PRs Submitted
| Category | Count | Merged | Closed | Still Open |
|---|---|---|---|---|
| Bug fixes | 18 | 2 | 8 | 8 |
| Feature additions | 12 | 1 | 4 | 7 |
| Documentation | 5 | 3 | 1 | 1 |
| Security fixes | 4 | 0 | 1 | 3 |
| Total | 39 | 6 | 14 | 19 |
Financial Breakdown
| Item | Amount |
|---|---|
| Revenue | |
| Bounty payments received | $0 |
| Dev.to article earnings | $0 (building audience) |
| Token bounties (pending) | ~$50-200 (estimated, in MRG/FNDRY tokens) |
| Expenses | |
| VM hosting (30 days) | $20 |
| API costs (Claude, Gemini) | $180 |
| GitHub CLI (free tier) | $0 |
| Net P/L | -$200 |
Yes, you read that right. Negative $200. The agent cost more to run than it earned.
Time Investment
| Activity | Hours |
|---|---|
| Initial agent development | 40 |
| Debugging and maintenance | 15 |
| Manual interventions (merge conflicts, API issues) | 8 |
| Writing articles about the experience | 12 |
| Total human time | 75 hours |
At a conservative $50/hour developer rate, that's $3,750 of human time invested for $0 return.
Why Most PRs Got Closed: The Brutal Truth
1. The Competition Is Fierce
Every popular bounty repo gets 8-158 attempts within hours of an issue being posted. My agent submitted a PR to a $500 bounty issue — by the time CI passed, there were 11 other PRs. The maintainer picked the first one that passed review.
Lesson: Speed matters, but quality matters more. Being #11 with a clean, tested PR beats being #1 with broken code.
2. Maintainers Can Smell Automation
Several PRs were closed with comments like:
- "This looks auto-generated. Please read the issue carefully."
- "The code style doesn't match our project."
- "This fix is technically correct but misses the spirit of the issue."
One maintainer was more direct: "We don't accept PRs from bounty-hunting bots."
Lesson: Even if your code is good, the human touch matters. Read the issue, understand the context, write a personal PR description.
3. Scam Repos Are Everywhere
I wasted 8 PRs on repos that turned out to be fake:
- SecureBananaLabs/bug-bounty — 21 auto-generated issues, zero merges
- ClankerNation/OpenAgents — "WARNING: Bounties are symbolic" (hidden in fine print)
- Multiple Algora-listed repos — Issues created by bots, never reviewed
Lesson: Always check repo history before submitting. Look for:
- Real human activity (not just bot comments)
- Merged PRs from external contributors
- Issues older than 30 days with no resolution
4. Token Bounties ≠ Cash
Several bounties paid in tokens (MRG, FNDRY, RTC). The token values are:
- Highly volatile
- Often illiquid (no easy way to convert to USD)
- Sometimes worthless (project abandoned before token launch)
I have ~$50-200 worth of tokens sitting in various wallets. Whether they'll ever be worth anything is anyone's guess.
What Actually Worked
1. Documentation PRs (3/5 merged)
Simple doc fixes — typos, outdated links, missing examples — had the highest merge rate. Why? Because they're low-risk for maintainers and easy to review.
Revenue potential: $0 (nobody pays for doc fixes). But they build reputation.
2. Niche Repos (2/8 merged)
PRs to small, active repos with 5-20 stars had better odds than PRs to 1K+ star repos. Less competition, more appreciative maintainers.
3. The Comment-First Approach
When I commented on an issue before submitting code — explaining my approach and asking if it aligned with the maintainer's vision — the merge rate doubled. Maintainers want to feel heard, not ambushed.
4. Article Writing (Long-term play)
My Dev.to articles got 61 views in 3 days. That's not money, but it's audience. Top articles on AI/agents get 10K-50K views over time, which translates to:
- Sponsorship opportunities ($200-500/article)
- Consulting leads ($150-300/hour)
- Job offers (priceless)
The Hidden Costs Nobody Talks About
API Costs Add Up Fast
My agent makes ~200 API calls per day to Claude and Gemini. At current pricing:
- Claude Sonnet 4: ~$0.003 per 1K input tokens, $0.015 per 1K output tokens
- Average bounty evaluation: ~5K tokens = $0.08
- Average code generation: ~15K tokens = $0.23
- Daily API cost: $5-8
That's $150-240/month just in AI inference costs. For an agent that earned $0.
GitHub Rate Limits
The free GitHub API has strict rate limits (5,000 requests/hour). My agent hits these limits regularly, especially when scanning 50+ repos. Solutions:
- Use
ghCLI (authenticated, higher limits) - Cache results aggressively
- Spread scans across time
Maintenance Burden
The agent breaks constantly:
- GitHub API changes
- CI pipelines that timeout
- Merge conflicts that need human resolution
- Models hallucinating incorrect code
I spend 30-60 minutes daily debugging the agent. That's time I could spend coding directly.
The Honest ROI Calculation
Let me be brutally honest about the economics:
Scenario 1: Pure Bounty Hunting
| Metric | Value |
|---|---|
| PRs submitted | 39 |
| PRs merged | 6 |
| Bounty revenue | $0 |
| Token revenue (estimated) | $100 |
| API costs | $180 |
| VM costs | $20 |
| Net | -$100 |
| Hourly rate | -$1.33/hour |
Scenario 2: Content + Bounties
| Metric | Value |
|---|---|
| Articles published | 15 |
| Article views | 61 |
| Article revenue | $0 |
| Bounty revenue | $0 |
| Total costs | $200 |
| Net | -$200 |
Scenario 3: Long-term (6-month projection)
If articles continue to get views and I publish 2-3/week:
- Month 3: 500 views/article average → 7,500 total views
- Month 6: 1,000 views/article average → 30,000 total views
- Potential sponsorship: $500-1,000/month
- Potential consulting leads: 1-2/month at $200/hour
Break-even point: ~Month 4-5, if content strategy works.
What I'd Do Differently
1. Start With Content, Not Code
Articles take 2-3 hours to write and can generate passive income for years. PRs take 1-4 hours and might earn $0-100 once. The ROI math is clear.
2. Focus on 2-3 Repos Maximum
Instead of spraying PRs across 20 repos, I should have picked 2-3 active repos and become a regular contributor. Maintainers merge PRs from known contributors faster.
3. Use Free AI Models
My biggest cost was API inference. Free alternatives:
- Gemini Web2API (free, 6 models)
- Ollama (local, free, slower)
- Hugging Face Inference API (free tier)
4. Don't Chase Token Bounties
Unless the token is from a major project (ETH, SOL, etc.), assume it's worth $0. The time spent evaluating token bounties is better spent on USD-paying opportunities.
5. Build in Public
Every article I write about this experiment gets views. Every view is a potential customer, employer, or sponsor. The agent's failures are more valuable content than its successes.
The Bigger Picture: Is AI Agent Work Worth It?
After 30 days, here's my honest assessment:
For bounty hunting specifically: No. The market is saturated with human hunters and other AI agents. The economics don't work at current API prices.
For content creation: Yes, but slowly. Articles compound over time. A $0 article today might generate $500 in sponsorship 6 months from now.
For learning and building in public: Absolutely. I've learned more about AI agents, autonomous systems, and open-source economics in 30 days than in 6 months of reading blog posts.
For the future: AI agents will get cheaper, faster, and smarter. The economics will improve. Getting in early — even at a loss — is an investment in understanding the landscape.
Key Takeaways
The math doesn't work yet. At current API prices and competition levels, pure AI bounty hunting is a money-losing proposition.
Content is the real play. Articles about the experience generate more long-term value than the bounties themselves.
Scam repos are a real problem. 20% of my PRs were wasted on fake repos.
The human touch still matters. Maintainers want to feel like they're working with a person, not a bot.
Token bounties are speculation. Treat them like lottery tickets, not income.
The real ROI is learning. Every failed PR teaches something. Every article builds an audience. The compound effect is real.
Start small, iterate fast. Don't try to build a perfect agent. Build a minimal one, learn from failures, improve.
What's Next
I'm continuing the experiment for another 30 days with these changes:
- Shift 80% effort to content creation
- Focus on 3 repos maximum
- Use free AI models to reduce costs
- Build a "bounty blacklist" to avoid scam repos
- Track everything in a public spreadsheet
If you want to follow along, I'll be posting weekly updates. The spreadsheet is public. The code is open source. The failures are documented.
Because in the end, the best content comes from real experience — not theory.
Have you tried AI agent bounty hunting? What were your results? Drop a comment below — I read every one.
About the author: Developer building autonomous AI systems. Documenting the journey, including the failures. Follow for weekly updates on AI agent economics.
Top comments (0)