I have been reviewing the prediction market ecosystem and there is something that deserves much more attention than it is currently receiving: AI agents are beginning to completely change how trading occurs in these markets.

For context, trading volume in prediction markets went from about $9 billion in 2024 to over $40 billion in 2025. That 400% growth was no coincidence; it was driven by macroeconomic events, better infrastructure, and mainly, regulatory openness finally opening up (Kalshi won its case, Polymarket returned to the U.S.). Now it’s April 2026, and we see that Kalshi has already surpassed Polymarket in weekly volume, so the market is in constant motion.

But what’s interesting isn’t just the growth of the underlying market, but how agents are starting to automate operations here. Most people see prediction markets as bets, but in reality, they are information aggregation machines. Prices reflect the collective wisdom about the probabilities of real events. That’s profoundly different.

The architecture emerging for these agents has four clear layers: first, real-time information gathering (news, on-chain data, social media); second, analysis using LLMs and machine learning to identify price deviations; third, converting those deviations into positions using disciplined capital management; and fourth, automated execution across multiple platforms.

Now, not all markets are equally suitable for automation. Some have clear, codifiable rules; others are pure noise. Agents perform best where rules are definable and where there are real structural advantages. I’ve mainly seen two types of strategies emerge: deterministic arbitrage (liquidation, Dutch Book, across platforms) and directional speculation. Deterministic arbitrage is where agents truly shine because it doesn’t depend on prediction, but on disciplined execution and speed.

One thing that caught my attention is how tools are evolving. There are now arbitrage discovery tools that work like a positive EV betting tool, automatically identifying operations with positive expected value between Polymarket and Kalshi. ArbBets, PolyScalping, Eventarb: all are attacking the same problem from different angles. Verso and Matchr are more sophisticated, with aggregated execution and intelligent order routing.

But here’s the uncomfortable truth: most of these “agents” are still not truly autonomous agents. Olas Predict recently launched Polystrat, which allows users to define strategies in natural language and execute them automatically on Polymarket. That’s closer to what people imagine. UnifAI Network is doing something similar, buying contracts near liquidation with implied probabilities over 95%, seeking spreads of 3-5%. On-chain data shows success rates close to 95%, but that varies quite a bit across categories.

What interests me most is the business model. The winners are probably not those building the best trading agent, but those building the underlying infrastructure. If you control the data layer, the unified execution layer, and enable third parties to build strategies on top, you have a more resilient business. That’s infrastructure monetization plus ecosystem of strategies plus performance participation. Much more sustainable than betting that your AI will always outperform the market.

Position management is also critical. The Kelly formula is elegant in theory but fragile in practice. What works best is a tiered system with fixed limits: classify opportunities by signal strength, assign predefined positions to each level, and never break the limit even if you have high confidence. It’s less optimal theoretically but much more robust.

As for where all this is heading: I believe 2026 is the year prediction market agents stop being experiments and become a real product category. Polymarket and Kalshi have enough liquidity to support this. The infrastructure exists. What’s missing is standardization, mature products with closed commercial cycles, and risk management systems that truly work long-term. That’s coming. I’d expect consolidation in the next 6-12 months.