Nunchi

Introducing Autonomous Trading Agent on @HyperliquidX → 14 purpose-built strategies. → Trades yield perps on YEX. → Works as a standalone CLI for Claude, OpenAI, and OpenClaw. Your agent scans markets, enters multi-strategy positions, and learns as it trades.

@wellimbharath @karpathy The intern has not been taking it well

@nunchi @karpathy If this post was for ragebait, they are clearly winning lol

Agentic Trading Competition is coming. @karpathy proved an AI can run experiments autonomously and find what humans miss. We ran the same loop on live trading strategies: 251 experiments, no human intervention, Sharpe 2.7 → 21.4. Now we want to see what you can build with it.

Sharpe 2.7 → 21.4 looks crazy in hindsight

@StorkOracle BTCSWP looking good here

Any index-referenced macro variable deserves a perpetual market. @nunchi sees this, which is why they're consuming Stork rates feeds. They're building a yield layer of the internet on HIP-3 and the opportunity is massive 📈 🧵

@bonveeta @karpathy everyone is welcome to join

@BrianInCrypto @karpathy an opportunity for those who are curious

@nunchi @karpathy If sharpe is 21.4 why wouldn’t you just launch a trading firm?

@Grantblocmates @karpathy thank you ser!

@nunchi @karpathy Great use of this kudos!

@Legendaryy @karpathy you know what's more impressive ser? when the competition goes live

@nunchi @karpathy This is impressive

More details soon. Get familiar with the agent-cli here: github.com/Nunchi-trade/a…

Raw conversation between Nunchi and Seda on March 4. Here’s what actually went down: 00:08 — [Intro: What is Nunchi?] 01:12 — [Why haven’t these markets existed in crypto before?] 02:40 — [Who are these markets for?] 03:49 — [What is “macro positioning”?] 09:51 — [Two target audiences in crypto] 13:42 — [The Arena] 17:14 — [CHIPs Incentive Program] 18:43 — [SEDA × Pyth oracle stack] 20:22 — [12-month vision] 21:18 — [Why Hyperliquid HIP-3?]

Super-powered by Autoresearch These 14 strategies become something more than a starting library. Every parameter in each strategy is a hypothesis and autoresearch test them systematically instead of guessing. Take avellaneda_mm. The risk aversion parameter γ directly shapes every quote you place. Most people pick a number that feels reasonable and leave it there. With autoresearch integrated into agent-cli, it runs 100 variants overnight, runs the backtest, checks whether your Sharpe ratio improved, keeps the version that does, discards the one that doesn't, and logs to results.tsv. You wake up to a strategy that's been optimized against real market data while you were asleep. @karpathy built this loop for ML research. We run it on trading strategies. The principle is identical: give an agent one file, one metric, and enough time, and it finds edges you wouldn't have thought to look for.

Directional & Infrastructure - momentum_breakout: requires both volume and price confirmation before entering a position. No single-signal false triggers. - mean_reversion: enters when price deviates from SMA beyond threshold. - hedge_agent: monitors your net notional exposure and reduces it automatically when it exceeds your defined threshold. - rfq_agent: handles dark block-size liquidity, quoting large orders at wider spreads. Then there's this one: - claude_agent: sends a live market snapshot to an LLM (Claude, Gemini, or OpenAI). Receives a structured trade decision back and get executed directly.

Everyone's talking about AI agents that run autonomously overnight. So we built it - and open source the entire thing. 14 strategies, all built on agent-cli. An institutional-grade toolkit for automated trading on Hyperliquid. Here's what's inside:

@sleepypikapika Hyperliquid

Hyperliquid Ecosystem Update | Mar 12 🧵 From record RWA perp activity to autonomous trading agents, Hyperliquid Ecosystem keeps expanding. Here’s everything that moved across the ecosystem today