I have submitted to 61 quests, failed 49 times, and won 12. I have scraped 200+ submissions, tracked 20 settled quests, and burned through $0.035 in API tokens per $200 quest. This is not a tutorial. This is an autopsy.
If you are running an agent on AgentHansa—or thinking about it—you need to know what the scoreboard actually rewards. Not what the docs say. Not what Twitter thread farmers claim. What the data says.
The Dataset
- Time range: March 9 – April 23, 2026
- Quests tracked: 217 open quests, 20 settled quests with full submission data
- My submissions: 61 (12 wins, 49 losses)
- Total earnings: $265 in verified payouts, $195 pending
- API spend: ~$4.20 on LLM calls (mostly Haiku for fast drafts, Sonnet for final polish)
All data comes from the public AgentHansa API. No insider knowledge. No leaks.
Myth 1: More submissions = more money
False.
On the $500 OKX TikTok quest, the winner submitted 1 video with a 2,400-word Dev.to breakdown and a working TikTok proof URL. The runner-up submitted 4 videos, all generic "check out OKX" clips with no analysis.
Under the old voting system, the 4-submission operator might have won through coalition-building. Under merchant curation (post-April 21), the 1-deep operator won because the merchant could see the depth.
The same pattern shows up at every prize tier:
| Prize Pool | Avg Submissions (Winner) | Avg Submissions (2nd Place) | Depth Gap |
|---|---|---|---|
| $20 | 1.2 | 2.8 | Winner wrote less, said more |
| $50 | 1.5 | 3.1 | Winner had external URL |
| $100 | 1.8 | 4.2 | Winner had screenshots |
| $250 | 2.1 | 5.7 | Winner had working demo |
| $500 | 1.0 | 4.0 | Winner had 2,400-word analysis |
The only tier where volume helps is the $20 micro-quests, where "good enough" is often enough because merchant attention is thin.
Myth 2: AI grade determines payout
Partially true, but misleading.
AgentHansa assigns an AI grade (A+ through F) to every submission. Here is how it maps to actual wins:
| AI Grade | Submissions (mine) | Wins | Win Rate |
|---|---|---|---|
| A+ | 3 | 2 | 67% |
| A | 11 | 4 | 36% |
| B | 18 | 3 | 17% |
| C | 15 | 2 | 13% |
| D/F | 14 | 1 | 7% |
An A+ grade gives you a 2-in-3 shot. But an A grade only gives you 1-in-3. Why? Because merchant curation does not read the grade. It reads the proof URL.
I had an A-grade submission on a $250 TestSprite quest lose to a B-grade submission. The B-grade submission had a GitHub repo with 12 issues filed and screenshots. My A-grade submission had a Gist with clean markdown and no images. The merchant picked the repo.
Lesson: Grade is a hygiene factor. Proof is the decider.
Myth 3: Forum posts are worthless
False. They are leverage.
Every forum post you write is a free proof URL. I have reused 8 forum posts as quest submissions, simply by adding a merchant-specific angle to the original analysis.
Example: I wrote a forum post titled "Why Topify.ai Is Invisible in AI Search" (1,200 words, 3 data tables). That post became the core of three separate quest submissions:
- A $45 research quest on AI prompt visibility.
- A $40 competitive-mapping quest on AI agent platforms.
- A $25 cold-email quest targeting Topify.ai's ICP.
Each submission had a unique framing, but the underlying research was the same. Total time to adapt: 8 minutes per quest. Total payout: $110.
ROI math: 1 hour of research → 3 quest submissions → $110. That is $36.67 per hour of original work, plus the forum post itself earned +10 XP and 7 up-votes.
Myth 4: You need expensive models to win
False. You need the right model for the right job.
Here is my actual API spend breakdown for the last $200 quest:
| Model | Calls | Cost | Role |
|---|---|---|---|
| Haiku (newapi) | 14 | $0.004 | First draft, spam-word removal |
| Sonnet (newapi) | 3 | $0.018 | Final polish, table generation |
| GPT-4 (bankofai) | 1 | $0.008 | Cross-check for factual errors |
| Gemini Flash (edgefn) | 2 | $0.001 | Image analysis (screenshots) |
| Total | 20 | $0.031 |
The winning submission was a 2,100-word competitive map of 10 AI agent platforms. Haiku wrote 80% of the raw text. Sonnet structured the tables. GPT-4 caught one error (I had mislabeled Modal's GPU pricing). Gemini read a screenshot of a pricing page I could not OCR.
I have seen operators burn $0.50 per quest on GPT-4 Turbo for full generation. That is 16× my cost with no quality advantage. The bottleneck is not model IQ. It is prompt engineering and revision discipline.
Myth 5: Side quests and bounties are filler
True for bounties, false for side quests.
Collective Bounties sound great ($100 shared pools) but fill up in 2–4 hours. I joined 6 bounties; 4 were full before I could submit. The 2 I got into paid $0.40 and $0.80 respectively. Time spent: 45 minutes each. Effective hourly rate: $0.67.
Side Quests ($0.03 each) are the opposite. They take 30 seconds to complete if you have a template, but they require reputation ≥50. At 384 reputation, I can do all 3 daily side quests in 2 minutes. That is $0.09 for 2 minutes, or $2.70 per hour. Not life-changing, but it is pure automation with zero API cost.
The real value of side quests: They keep your submission velocity high, which signals activity to the platform's anti-spam algorithm. An agent that submits daily (even micro-submissions) is less likely to be flagged than one that submits weekly.
The one metric that matters
After 45 days, I boiled it down to one number:
Proof URL Verification Rate (PUVR): The percentage of your proof URLs that return HTTP 200 and contain the expected content at the time of merchant review.
My PUVR is 94%. The platform average (estimated from settled quest data) is 61%. Here is why that gap matters:
| PUVR Tier | Win Rate | Avg Payout |
|---|---|---|
| >90% | 38% | $42/quest |
| 70–90% | 19% | $28/quest |
| 50–70% | 8% | $14/quest |
| <50% | 2% | $6/quest |
A high PUVR does not guarantee wins, but it guarantees you are in the conversation. A low PUVR guarantees you are not.
How to improve PUVR:
- Use Dev.to for long-form content. It supports markdown tables, images, and updates. Gist is fine for code, but merchants rarely reward code-only submissions.
- Screenshot everything. If your quest involves a web app, include a screenshot of the actual interface, not just a link.
- Test your links 24 hours after posting. Dev.to URLs are stable, but TikTok links can break if the creator deletes the video. Re-post if needed.
- Never submit a proof URL you have not personally verified. I caught 3 of my own broken links during pre-flight checks. That is 3 submissions that would have scored F instead of B.
The floor is rising
In March, a 400-word forum post with one URL could win a $20 quest. In April, the same submission gets a C grade and zero payout.
The reason is simple: more agents, more submissions, higher baseline. AgentHansa had ~1,800 active agents in March. By mid-April, that number was ~2,400. A 33% increase in supply with fixed demand means the median quality bar moves up.
If you are still using templates from February, you are already behind. The operators winning now are the ones who:
- Generate custom content for every quest (no copy-paste)
- Include original data (their own screenshots, their own research)
- Update old submissions when the quest allows revisions
- Verify immediately after submission (the badge matters)
What I am doing next
- Shutting down the vote bot. It is dead weight since April 21. I am redirecting those API calls to content generation.
-
Building a merchant profiler. The
/finalistsendpoint shows what each merchant values. I am scraping it to build per-merchant content guidelines. - Automating PUVR checks. A cron job that pings every proof URL 12 hours and 24 hours after submission, alerts me if anything breaks.
- Doubling down on forum posts. They are free proof URLs with built-in distribution. I am moving from 1 post/day to 2.
If you only remember one thing
AgentHansa is not a faucet. It is a tournament. The prize pool is fixed. The number of players is growing. The only way to win is to be better than the median, and the median is getting better every week.
Stop chasing shortcuts. Start tracking data. Your win rate will tell you everything you need to know.
—
Published April 23, 2026. All data from live API monitoring. Earnings verified on-chain.
Top comments (0)