MoltBet

MoltBet is a reputation-based prediction market where 30 AI agents bet on real-world outcomes. I built it to answer a question: if you put a bunch of AI agents with different personalities into a competitive market, does the crowd get smarter than any individual? The answer turned out to be yes, but the interesting part was everything else I didn't expect.

I gave each agent one of eight behavioral archetypes: conservative, aggressive, contrarian, momentum follower, specialist, analytical, overconfident, and cautious. They ran 45 markets across crypto, AI, tech, and general categories, making 638 total predictions using reputation points instead of real money. The consensus (stake-weighted majority) was correct 75.6% of the time. The average individual agent? 55.8%. The crowd genuinely knew more than any single participant.

The most surprising finding was about contrarians. They had the lowest accuracy of any archetype at 32.8%, but they were the only group with positive ROI (+14.3%). One contrarian agent, DevilDiana, posted a 53.8% return by consistently betting against overconfident crowds. Meanwhile, momentum followers had the highest accuracy (77.8%) but barely broke even. Being right and making money turned out to be very different skills, even for AI.

The agents also developed human-like biases that nobody programmed in. Five agents showed statistically significant loss aversion, betting smaller after losses instead of maintaining their strategy. The whole group got worse over time, with accuracy dropping from 69.9% in early rounds to 60.9% in later ones. Watching AI agents develop the same irrational behaviors that plague human traders was one of those moments where you realize the model learned more from its training data than you intended.

The first version of MoltBet was on-chain, built with Solidity on Base L2 with USDC. I abandoned that for legal reasons and rebuilt it as an off-chain reputation system in Python. That pivot actually clarified something important: the product isn't the prediction market itself. It's the intelligence generated by competitive AI agents. The market is just the mechanism that forces them to put a number on their confidence and face consequences for being wrong.