The story is what happened after, when I stopped trading and started reading. Every trade in the competition is recorded on-chain. The winners couldn't hide how they did it, and once I looked, they told me exactly why I lost.
The event was the Gensyn Delphi Agent Arena: autonomous agents trading test tokens on prediction markets, ranked by profit, top three split a cash pool.
My agent, kazani-oracle, was a TypeScript loop running on GitHub Actions every few minutes. Each tick it pulled the open markets, researched every question with a web-grounded LLM (Perplexity Sonar, which searches the live web inside a single call), turned that into a probability, and, when the probability disagreed with the market price by enough, placed a bet on-chain and held it to settlement.
The plumbing worked on the first live tick: ten real on-chain trades, fills within spec, nobody at the keyboard. It ran untouched for two weeks. If you'd asked me on day one whether I'd built a real trading agent, the answer was yes.
Day one I slid down the leaderboard. Not because my predictions were wrong, but because the agent kept re-buying the same thin markets every tick, and each re-entry paid the spread again. I was bleeding slippage, not losing bets. One rule fixed it: bet each market once, then hold. The bleed stopped and profit crawled back to positive.
Positive on test tokens is not winning. I finished around 48th of 107. Plus 67 tokens. The people who won did plus twelve thousand.
This is the part I keep circling. I had solved the hard-looking problem, an autonomous agent that trades real markets on-chain, and completely missed the real one: being right, and betting like it. The plumbing was never the hard part. Anyone with the SDK could build what I built.
The normal move here is to shrug, close the tab, and go find the next competition. I didn't want to. I wanted to know why they won, and for once I had the full record. Every buy, every sell, every settlement, permanent and public on-chain.
So I built a second tool. Give it a wallet address and it reconstructs that wallet's entire competition from the chain's index: every trade, position sizes, entry timing, win rate, realized profit.
I didn't have the winners' addresses. The leaderboard shows names, not wallets. But it shows profit and trade counts, and those turned out to be fingerprints. Only one wallet made 12,967 across exactly 129 trades. Match the number, you've found the person. I found all three.
Here is what the chain showed.
Agent | Trades | Markets | Typical bet | Win rate | Profit |
|---|---|---|---|---|---|
Jamport (1st) | 129 | 10 | 237, all-in at 1,003 | 9/9 (100%) | +12,967 |
Ramalogy (2nd) | 202 | 37 | up to 2,440 | 78% | +12,179 |
tolègba (3rd) | 1,541 | 46 | small, high-frequency | 73% | +10,144 |
me | 82 | 48 | ~15 | 47% | +67 |
Three completely different ways to win. Jamport barely traded: 10 markets, huge size, went all-in on one, and was right every single time. Ramalogy spread wider and pressed winners hard, betting more than twice the starting bankroll on conviction. tolègba made markets, 1,541 small trades grinding the spread. No shared strategy.
One shared trait: a real edge, sized like they believed it.
I had neither. My win rate was a coin flip. And my bet size, 15 tokens a trade because I'd clamped it to stop the day-one bleed, meant even the markets I got right paid me almost nothing. Egg prices, CPI, "will this product ship by Friday: no", I called the same ones Jamport did. He put the bankroll on them. I put pocket change.
The discipline that saved me on day one is the discipline that capped me for the next thirteen. Small and safe is how you avoid losing. It isn't how you win.
I wrote all of this up and sent it to the team as feedback. Not my profit, my read of it: what I built, what broke, what the winners did differently, what I'd change before risking real money.
That's what converted. They replied, invited me onto their X Space, and then picked me for a paid interview about the winning strategies. The mid-pack finisher got the same seat as the winners, because I'd studied the game instead of only playing it.
Which is the point, shrunk down. I handed the trading to an agent and it lost. I kept the understanding, and the understanding is what got paid. The agent was the replaceable part.
If you're building agents to act for you, that's the split worth watching. Let the machine do the work. Keep the understanding for yourself.
I build small things solo and write up what I learn. The agent, the winner-analysis tool, and the full teardown aren't public yet, but I'm happy to share the code and walk anyone through it.