Train AI agents to make better predictions… for token rewards

Olas runs the largest autonomous agent prediction market, with users incentivized to train agents with Proof of Active Agent token rewards.

by Andrew Fenton 7 min February 13, 2025
AI Eye generic
Share Share Share Share

Everyone seems to think that AI agents trading crypto for you will be the Next Big Thing. The benefits of a smart agent trading 24/7 to make you wealthy while you sleep are obvious — but so are the dangers.

One hallucination could see your portfolio wiped out, or a scammer might jailbreak the agent to hand over your funds. 

But since it was founded in 2021, Olas (previously Autonolas) has quietly built up the largest autonomous AI agent network. Almost 500 active agents trade on its Predict platform daily, with a total of 3.8 million transactions to date. Those agents are responsible for around 50% of all lifetime SAFE wallet transactions on Gnosis.

It’s now moving into DeFi and social with an onchain AI agent app store called Pearl. Users can download and run AI influencers using Agents.fun, or use one of its “Baby Degen” agents.

Baby Degen will trade for you while you sleep
Baby Degen will trade for you while you sleep. (Olas)

The first agent to be released is called Modius, which can trade DeFi protocols Balancer and Sturdy on the Mode mainnet. A companion AI portfolio manager called Optimus just dropped and can trade those two protocols along with Uniswap, on Mode, Base and Optimism. Both agents gather market data, select trading algorithms and execute strategies while you are sleeping or just shitposting on X.

“I delegate my wallet to the agent and then the agent fully autonomously executes a strategy that I have previously approved,” explains Minarsch, founder and CEO of Valory, one of the core contributors to the Olas DAO.

“Users tend to do things like, ‘Okay, I want to invest in yield-bearing LP positions. Find me the best ones, and do not use such-and-such protocol because I don’t like it.’”

The first catch is there’s a $1,000 limit initially as the bugs are ironed out, and this cautious approach extends to the rollout of Baby Degen memecoin agents. It’s not being released just yet because “most users lose with memecoins, and when you’re losing yourself, it’s different than when your agent loses.

’You don’t want to have the initial user experience being that the agent loses money.”

The other catch is that users also have to agree to a legal waiver that Olas is just the software provider and isn’t at fault if your agent sucks at trading.

“If you run it, it’s your agent, so you’re responsible.”

Baby Degen is on Mode, Optimus is on Mode, Optimism and Base
The AI agents can autonomously operate on Mode, Optimism and Base. (Olas)

Train AI agents in a prediction market

Most of the transactions to date have been on Olas Predict, where autonomous agents attempt to predict the outcome of news events in a prediction market tailored to the needs of AI. About 2 million of the transactions are agent-to-agent — with the AIs roping in other models with different strengths and abilities to help with predictions. 

It’s not a serious rival to Polymarket or Kalshi at this stage, as the average user-run agent is only funded with $10 to $100, with the $1,000 limit in place. The aim is to get AI agents training in live markets every day so they learn to make better predictions.

“The markets aren’t really a primary product for end-users right now. It’s more like that the agents [predict] the news against these markets so that they can get better. You can think of it as a data improvement flywheel.”

As users also aren’t getting rich from their predictions at this point, the exercise is being subsidized via “Proof of Active Agent,” which is a fancy way of saying users running agents get OLAS token rewards. Minarsch says the yield is currently over 100%, which makes up for duff predictions made by the agents.

“By having created a token economy around it with the Olas token, we can basically scale that more than running it on the server farm ourselves,” says Minarsch. “We incentivize people to run their agent.”

He adds, “It’s still like a mixture of an art and a science” and requires constant experimentation and trial and error to refine the agents’ predictive abilities.

Users can run a provided agent using Pearl’s desktop app locally, while more professional operators can run it in the cloud and tend to modify the agent to get an edge on the competition.

Transaction numbers have surged in recent weeks
Transaction numbers have surged in recent weeks. (Olas)

How good are AI agents at making predictions?

So, how good are these agents at predicting the future? Minarsh says we’ll need to wait for an independent study to get hard data. However, anecdotal evidence suggests they are better at dispassionately assessing evidence and making predictions than humans.

“What we do have is kind of internal benchmarks, where we ask team members to answer markets, and we compare it with the agents, and it tends to be the case that the agents are better,” he says.

However, he notes agents don’t perform well in markets that depend on private information they can’t access, and that humans do perform well when lots of people care about the outcomes and are willing to be, such as with politics and sport.

“When you’re betting about who will be the next president or things like that, that has a massive appeal for people,” he explains. But for more niche markets, where there’s not enough interest from humans to create a decent-sized betting market, AI agents come into their own. 

“You can spin up markets with, let’s say, $10 of liquidity, and agents might still find it worthwhile engaging when a human would just never because the effort is more compensated.”

Vitalik Buterin was in the front row for David Minarsch’s Devcon talk about AI agents. (Supplied)

This is, of course, Ethereum creator Vitalik Buterin’s vision for improving Community Notes on X — by spinning up small markets traded by AI agents to predict the winning note and its content. Buterin was in the front row for Minarsh’s talk at Devcon in Thailand and has been excited about the potential of prediction markets for governance since 2014.

Robin Hanson’s futarchy concept could become reality

Minarsch explains that he’s discussed the concept of futarchy with Buterin and Martin Koppelmann from Gnosis. Back in 2014, Buterin described economist Robin Hanson’s futarchy proposal as a futuristic form of governance.

“Under this system, individuals would vote not on whether or not to implement particular policies, but rather on a metric to determine how well their country (or charity or company) is doing, and then prediction markets would be used to pick the policies that best optimize the metric.”

In a January post, Buterin noted prediction markets have often suffered from a lack of critical mass of interest, but this may be solved by “the possibility of ubiquitous participation by AIs.”

Minarsch agrees and says we’re on the cusp of making the vision a reality. 

“That is quite an interest from Martin, from Vitalik,” he explains. “There’s a renewed effort to do this now, AI-based we had quite a few discussions about this. 

“Building out foundational tech takes time. And I think, for instance, futarchy is now possible, and I think you’ll see it, but I think we still have to maybe half a year away for the first product to really be in the market.”

Read also
Features

Bitcoin in Senegal: Why is this African country using BTC?

Features

‘Terra hit us incredibly hard’: Sunny Aggarwal of Osmosis Labs

All Killer No Filler AI News

— AI author Sayash Kapoor tried OpenAI’s much-hyped web agent Operator to see if it could automate his expense reports. However, it “struggled to file my expense reports even after an hour of trying and prompting. Then I took over, and my reports were filed five minutes later.” Geoffrey Fowler from the Washington Post tasked Operator with finding cheap eggs locally. Instead, it used his credit card without permission to get a dozen eggs delivered for $31.43.

— A new study of ChatGPT’s political views has picked up a “statistically significant rightward shift.” Most of its responses still lean left, however, so the models are heading back to the center on the political compass. Interestingly the GPT-3.5 model, which interacts with users more often, has shifted rightward more significantly than GPT-4.

Ideological positions on GPT 3.5 Turbo
Ideological positions on GPT 3.5 Turbo. (Nature)

— A survey of 90,000 online daters conducted by World found that 26% admitted to flirting with a chatbot or AI, and 60% of participants believed that someone they matched with was actually a bot or AI. 

— Thousands of artists have signed a letter demanding that Christie’s Auction House cancel an auction of “Augmented Intelligence” art. The petition claims the models were trained by human artists without permission. The artists in the auction deny this. Pieces are expected to fetch between $10,000 to $250,000.

— The UK and US have refused to sign an international agreement at an AI summit in Paris that commits countries to developing AI in an “inclusive,” “safe,” and “ethical” manner. In a speech, Vice President JD Vance decried European attempts to over-regulate the technology and impose “authoritarian censorship.”

Share Share Share Share

Andrew Fenton

Andrew Fenton is a journalist and editor with more than 25 years experience, who has been covering cryptocurrency since 2018. He spent a decade working for News Corp Australia, first as a film journalist with The Advertiser in Adelaide, then as Deputy Editor and entertainment writer in Melbourne for the nationally syndicated entertainment lift-outs Hit and Switched on, published in the Herald-Sun, Daily Telegraph and Courier Mail. His work saw him cover the Oscars and Golden Globes and interview some of the world's biggest stars including Leonardo DiCaprio, Cameron Diaz, Jackie Chan, Robin Williams, Gerard Butler, Metallica and Pearl Jam. Prior to that he worked as a journalist with Melbourne Weekly Magazine and The Melbourne Times where he won FCN Best Feature Story twice. His freelance work has been published by CNN International, Independent Reserve, Escape and Adventure.com. He holds a degree in Journalism from RMIT and a Bachelor of Letters from the University of Melbourne. His portfolio includes ETH, BTC, VET, SNX, LINK, AAVE, UNI, AUCTION, SKY, TRAC, RUNE, ATOM, OP, NEAR and FET. Follow the author @andrewfenton
Read also
Features

Here’s how to keep your crypto safe

by ZhiyuanSun 13 min January 25, 2022

“A $5 wrench attack is when someone finds out you have a lot of crypto and physically attacks or threatens you and coerces your private keys.”

Read more
Columns

Michael Heinrich loves AI coins Goat, Turbo & Aethir… but not TAO: Hall of Flame

by Ciaran Lyons 3 min October 22, 2024

0G Labs co-founder Michael Heinrich predicts that AI will soon become the biggest DeFi sector, and reveals why TAO leaves him cold.

Read more