AI agents used to trade crypto
A viral clip and posts say a Chinese developer’s tutorial accidentally revealed about $868k in profits from Claude agents that traded more than 28,000 BTC‑sized positions on Polymarket using an AI farm and 15‑minute intervals, and that clip garnered roughly 400k+ views (x.com) (x.com). The posts fueled community discussion about agentic trading experiments and the practical limits of automating high‑frequency decision flows with LLM agents (x.com) (x.com).
Prediction markets are online exchanges where traders buy “yes” or “no” shares on future events, and Polymarket offers application programming interfaces that let software place and manage those orders automatically. (docs.polymarket.com) That plumbing is what turned a recent viral clip into a broader argument over “agentic” trading: posts on X said a Chinese developer’s tutorial screen showed Claude-based agents making about $868,000 while trading more than 28,000 Bitcoin-sized positions on Polymarket in 15-minute cycles. The clip and follow-up posts together drew roughly hundreds of thousands of views, according to the posts themselves. (x.com 1) (x.com 2) Polymarket’s own documentation shows why the claim sounded plausible to traders. Its developer tools support market data access, order creation, order cancellation, batch submission, and real-time order updates, which are the basic parts needed for a bot to scan markets and trade without clicking through the website by hand. (docs.polymarket.com) (docs.polymarket.us) Anthropic has also been pushing Claude toward longer multi-step work. The company said Claude can use tools, write code, and in “computer use” mode interact with a screen by moving a cursor, clicking buttons, and typing, while warning that the feature was released as an experimental beta and could still be error-prone. (anthropic.com) That combination has produced a small wave of public Polymarket agent projects in 2026. GitHub repositories and blogs now show Claude-based systems that wake up on a schedule, call Polymarket tools, search the web for facts, and place trades with little or no manual intervention after setup. (github.com) (medium.com) The practical limit is speed. Polymarket’s trading stack supports streaming updates and batch order handling, but large language model agents still spend time reading prompts, calling tools, and checking results, which makes them better suited to slower event-driven bets than to the millisecond-style trading common in traditional high-frequency markets. That gap is reflected in Anthropic’s own description of computer use as “experimental” and “error-prone.” (docs.polymarket.us) (anthropic.com) Cost is another constraint. Anthropic’s current pricing pages list Claude Pro at $20 a month if billed monthly, while its application programming interface pricing for recent models starts at $1 per million input tokens for Haiku 4.5, $3 for Sonnet 4.6, and $5 for Opus 4.6, with much higher output-token prices, so frequent multi-agent runs can add up quickly if a bot is constantly researching and revising trades. (claude.com) (anthropic.com 1) (anthropic.com 2) (anthropic.com 3) The viral posts did not by themselves prove the full trading history, wallet ownership, or realized profit behind the screenshot, and independent verification would require matching the claimed account or wallet to on-chain or platform records. What the clip did show, and what the documentation now makes easier to understand, is that the tools to let language-model agents research, decide, and trade on prediction markets are already in public view. (x.com) (docs.polymarket.com) (anthropic.com)
