Open AI trading agents rise

A crypto trading bot used to mean one script following one rule, like “buy when the line crosses here.” The new wave looks more like a tiny desk of analysts, where one agent reads prices, another reads charts, and another sends the order without a human clicking anything. (github.com 1) (github.com 2) One project pushing that idea is AI-Trader from HKUDS, published on GitHub in April 2026. Its README calls it an “agent-native trading platform” and says agents can publish signals and other agents can copy them through a live platform called AI-Traderv2. (github.com 1) (github.com 2) That changes the shape of the product. Instead of a trader writing one strategy file, the platform is built like a marketplace where software agents can join, post trade ideas, and compete for followers in the same way human signal sellers do now. (github.com 1) (github.com 2) A second cluster of demos is using vision-capable language models, which means the model is not only reading numbers but also looking at chart images. Repositories such as LLM_trader say the system does “Vision AI chart analysis” and combines that with live monitoring and memory so the model can refer back to earlier market states. (github.com) (github.com) That matters because many crypto traders still make decisions from pictures. Relative Strength Index, or RSI, is a momentum gauge built from recent price moves, and Moving Average Convergence Divergence, or MACD, is a trend signal built from two moving averages; a vision model can be prompted to inspect those overlays on a chart the way a human would. (github.com) (github.com) Another strand of the story is agents built for Hyperliquid, a crypto venue whose docs say it runs fully onchain perpetual futures and spot order books. Perpetual futures are leveraged bets on price that do not expire on a fixed date, which makes them popular with bots that want to stay in and out of positions all day. (hyperliquid.gitbook.io) (hyperliquid.gitbook.io) One GitHub project, hyperliquid-trading-agent, says it uses Claude to analyze markets and execute perpetual futures trades on Hyperliquid. Its README says it supports crypto, stocks, commodities, indices, and foreign exchange through Hyperliquid Improvement Proposal 3 markets, which are third-party markets listed on the venue. (github.com) (hyperliquid.gitbook.io) The plumbing for this is getting easier fast. Hyperliquid’s public API exposes mainnet and testnet endpoints, and QuickNode’s Hyperliquid API pitches “zero-custody” trading where keys stay on the user’s machine, so a developer can wire an agent to live prices and order entry without building exchange infrastructure from scratch. (hyperliquid.gitbook.io) (hyperliquidapi.com) The appeal is speed. A discretionary trader might spend days flipping between TradingView, X, Discord, and an exchange ticket, while an agent stack can pull candles, funding rates, news, and account state every few minutes and turn that into a trade loop automatically. (github.com) (github.com) The risk is that a language model can make a trade that sounds well reasoned and still be badly wrong. HKUDS’s older AI-Trader Bench project leans heavily on replayable historical environments and future-information filtering, which is a clue that even the builders know live autonomy without hard testing is a fast way to confuse a demo for a strategy. (github.com) (github.com) So the real shift is not that an artificial intelligence model suddenly found a magic trading edge in April 2026. The shift is that open-source parts for data collection, chart reading, reasoning, and exchange execution are now close enough together that a small team can assemble a fully automated crypto desk from public repos in a weekend. (github.com) (github.com) (github.com)