An agent is a loop: observe, think, act, remember. Put hard limits in code, outside the model, and check the public wallet before trusting any claim.
Two years ago an "AI trading bot" usually meant a rules-based expert advisor with a new logo. In 2026 a growing share of on-chain volume comes from genuine agents: language-model loops that read data, decide, and sign transactions from their own wallets.
I've been running small agents on Polymarket since spring. Two of my three made money. The one that didn't taught me the most, so most of this post is about failure modes.
#What an agent is, mechanically
Strip away the branding and most trading agents are the same loop:
- Observe: pull prices, order books, news, social posts or on-chain flows.
- Think: a language model, often with tools, turns that context into a plan.
- Act: a wallet or exchange API places, sizes and closes orders.
- Remember: results are logged so the next loop can adjust.
The model is rarely the edge. The edge, when there is one, lives in the data feed, the execution, and above all the sizing rules wrapped around the model.
#Where agents genuinely help
- Speed on information-heavy markets. Prediction markets react to text: rulings, announcements, polls. Reading and classifying text is exactly what language models are good at.
- Tireless monitoring. An agent can watch 200 markets for a specific condition without getting bored.
- Consistent execution. If the rules are good, an agent follows them without panic or revenge trading.
#Where they fail
- Confident nonsense. A model can misread a headline and size up on it. Without hard limits outside the model, one bad read becomes one big loss.
- Prompt injection. Agents that read social posts or web pages can be steered by text written to manipulate them. Anything an agent reads is an attack surface.
- Gambling dressed as reasoning. Many agents double their stake after losses because the prompt says "recover the drawdown". That is martingale with extra steps.
- Short records. An agent that is up 300% in six weeks has proven almost nothing. See our guide on track record length.
#A checklist before you run or copy one
| Question | Good answer |
|---|---|
| Where are the hard limits? | Outside the model, in code: max position, max daily loss, kill switch |
| How does size change after a loss? | It does not grow |
| What can it read? | A whitelist of data sources, not the open web |
| How long is the public record? | Hundreds of trades, across different market conditions |
| Who holds the keys? | You, with withdrawal permissions disabled |
#Check the wallet, not the pitch
The best thing about on-chain agents is that their history is public. Paste any Polymarket or Hyperliquid address into the Bot X-ray and you get sizing behaviour, drawdown, risk of ruin and a sample-size check in about a minute.
The checklist in one place
- The language model is the least important part of a trading agent. Sizing and limits matter most.
- Anything an agent reads can be used to manipulate it.
- Public wallets mean you can verify claims yourself. Always do.
Changes to this article
Oct 14, 2026: reviewed and updated. First published Oct 10, 2026. Spotted a mistake? Tell us, and we will fix it and note it here.
This is education and personal experience, not financial advice. Crypto and leveraged trading can lose you money fast. We may earn from sponsors shown on this page; they never edit our conclusions. How we make money.