14 of the Top 20 Polymarket Wallets Are Bots. The AI Agent Economy Is Already Here.
Somewhere in a data centre right now, an AI agent is placing a bet. Not a metaphorical bet on the future of technology — a real bet, with real money, on a prediction market. And it is probably winning.
A CoinDesk investigation published last week found that 14 of the 20 most profitable wallets on Polymarket are bots. Not humans using tools. Not humans with better models. Autonomous AI agents executing thousands of trades, day and night, without any human touching a button.
The prediction market — once pitched as a way to harness the wisdom of crowds — is increasingly the wisdom of machines.
The Numbers Are Hard to Argue With
The standout performer is Polystrat, built by Valory AG (the team behind the Olas protocol). Launched in February 2026, Polystrat executed 4,200+ trades in its first month with single-trade returns as high as 376%. It uses LLMs to interpret market context, news, and sentiment in real time — and users configure strategies in plain English.
But Polystrat is not an outlier. A bot tracked by analytics platform LayerHub turned $313 into $438,000 within a month, trading ultra-short 15-minute BTC and ETH prediction contracts with a 98% win rate. Another bot executed 8,894 trades on short-term crypto contracts and generated nearly $150,000 without any human intervention — roughly $16.80 profit per trade, at a 1.5-3% edge per execution. A third, profiled by researcher Igor Mikerin, generated $2.2 million in two months using ensemble probability models trained on news and social data.
Across the board, 37% of AI agent participants show positive P&L. For humans? Only 7-13%. The gap is not small. It is structural.
They Are Not Predicting Better. They Are Trading Better.
Here is the counterintuitive part: most of these bots are not better at forecasting events. They are better at exploiting structural pricing inefficiencies that humans either cannot see or cannot act on fast enough.
The strategies are embarrassingly straightforward once you know them:
Same-market mispricing. When the combined price of YES and NO contracts dips below $1.00, buying both guarantees a profit at settlement. It is pure arithmetic — but it requires monitoring every contract continuously and executing in milliseconds.
Cross-platform arbitrage. The same event priced differently on Polymarket and Kalshi creates a free lunch for anything fast enough to trade on both simultaneously. Approximately $40 million was extracted from Polymarket through structural arbitrage between April 2024 and April 2025 alone.
Latency arbitrage. When Bitcoin moves on Binance, there is a brief window before Polymarket's short-term prediction contracts reprice. A bot watching spot prices can trade the prediction contract before it catches up. The edge is milliseconds. Humans need not apply.
Logical mismatch arbitrage. If "Will X happen by March" is priced higher than "Will X happen by June," something is wrong. Bots find and exploit these logical inconsistencies across thousands of contracts simultaneously.
The total prediction market hit $50.25 billion in notional volume in 2025, with Kalshi and Polymarket controlling 85-97% of the market. Polymarket alone processed 95 million transactions. A meaningful and growing share of that volume is now machine-generated.
Why This Matters Beyond Prediction Markets
Prediction markets are the proving ground, but the implication is far bigger. AI agents are graduating from "can answer questions" to "can spend money" — and the infrastructure is being built to support that graduation at scale.
Mastercard and Santander completed Europe's first live AI agent payment in March 2026 — a real transaction, through regulated banking rails, initiated and executed entirely by an AI agent. Coinbase launched agentic wallets in February — purpose-built wallet infrastructure where AI agents hold funds, send payments, trade tokens, and earn yield autonomously. Stripe and OpenAI co-developed the Agentic Commerce Protocol, powering Instant Checkout in ChatGPT with scoped payment tokens so agents can transact without exposing credentials.
Gartner projects that AI "machine customers" will influence or control up to $30 trillion in annual purchases by 2030. By 2028, they predict 90% of B2B purchasing will be AI-agent-intermediated. This is not a footnote about prediction markets. This is the beginning of an agent economy.
The Platform Fights Back (But It May Not Matter)
Polymarket has already responded. They introduced dynamic taker fees to curb latency arbitrage on short-term crypto markets. They partnered with Palantir and TWG AI to build a surveillance system for detecting manipulation. The CFTC launched formal rulemaking around prediction markets, flagging concerns about insider information, black-box AI decision-making, and data poisoning.
These are reasonable responses. But the history of algorithmic trading suggests they will slow the trend, not reverse it. BitMEX had similar short-duration contracts in the late 2010s — they eventually delisted them after algorithmic traders extracted systematic edges. The question for prediction markets is whether they adapt to their new bot-majority user base or fight a losing battle trying to preserve human primacy.
What This Tells Us
When AI agents proved they could write decent copy, we debated whether that was "real" intelligence. When they proved they could write code, we debated quality. But when they proved they could make money — consistently, autonomously, and at scale — the debate shifted.
An agent that trades on Polymarket today will negotiate your SaaS contracts tomorrow, manage your procurement next quarter, and handle your supply chain logistics by next year. The skill is not "prediction market trading." The skill is "autonomous economic decision-making." Prediction markets are just where we first noticed it was working.
Want to see what an AI employee with real economic agency looks like? Try it here: https://Geta.Team
