We run a small experimental trading desk on Base, an Ethereum layer 2. One AI call a day reads the market and proposes trades, and a separate program checks those proposals against hard rules before anything gets signed. Before we add to it, we wanted to know what the rest of the field has learned. Which agent designs work for crypto trading? Which strategies have evidence behind them? And where do “skills”, the folders of instructions and scripts that agents like Claude Code and Hermes load on demand, fit in?
We read the papers, the competition results and about twenty open-source repos. Here is what we found, with links so you can check our reading.
LLMs read context well and read charts badly
The most useful result came from the most public experiment. In October 2025 the lab nof1 gave six frontier models $10,000 each to trade perpetual futures on Hyperliquid for about two and a half weeks, with the same prompt and the same data. Four of the six lost money, and the two worst ended with about $4,200 and $4,100. Qwen3 Max won with roughly 22%. The founder's own caveat is the important line: LLMs don't handle numerical time series well, and that was the entire context the models got. Two weeks is also far too short to separate skill from luck, so we read it as a warning, not a ranking.
The careful version of that warning is a paper called FINSABER (arXiv 2505.07078, accepted at KDD 2026). Earlier studies that showed LLM trading strategies beating the market tested them over short windows and a handful of stocks. FINSABER reran those strategies over two decades and more than 100 symbols, with controls for survivorship and look-ahead bias. The advantage “deteriorates significantly.” The models were too cautious in bull markets and too aggressive in bear markets. The authors' advice is to focus on trend detection and regime-aware risk control rather than bigger agent frameworks.
So we stopped asking which model trades best. A better question is which part of the job a language model is actually good at. Our answer is the parts that involve reading: news, governance votes, token contracts, protocol risk, and explaining why a decision was made. The price signal itself should come from plain code.
The strategies with evidence are boring
Time-series momentum, meaning hold an asset while its trend is up and step aside when it isn't, has the most support in crypto. A 2025 working paper by Zarattini, Pagani and Barbon (Catching Crypto Trends) applied simple channel breakouts to the top 20 coins, sized each position by its volatility, and reported a Sharpe ratio above 1.5 after fees. It's a working paper, and an older study found some short-horizon momentum faded once Bitcoin futures launched, so we treat the exact numbers loosely. The shape of the idea has held up across several studies.
Two other strategies have decent support. One is rebalancing between a stable core and a few small satellite positions. The other is moving idle stablecoins to whichever lending market pays best, but only when the rate gap covers the cost of moving. That second one is what Giza's ARMA agent does on Base.
A few popular ideas look worse on inspection. Providing liquidity on Uniswap and leaving it alone lost money in many of the largest pools, because arbitrage losses outran the fees (Fritsch and Canidio). Funding-rate carry still works but has shrunk; Ethena's staked USDe yield fell to around 3.5% at one point in 2025. We found no evidence that an LLM making intraday calls on a small account earns anything after costs.
Three ways to arrange the agents
The best-known open-source design is TradingAgents from Tauric Research. It models a trading firm. Analyst agents cover fundamentals, sentiment, news and technicals. A bullish and a bearish researcher argue over the analysts' reports, a trader proposes a transaction, a risk team with aggressive, neutral and conservative voices debates it, and a portfolio manager approves or rejects. It's Apache-2.0, built on LangGraph, and works with most model providers.
Two ideas from it are worth stealing even if you never run the whole graph. After each decision's holding period ends, it settles the call against the actual return and writes a one-paragraph lesson that later runs read. It also splits work between a cheap model for the analysts and a strong model for the final decision. The full debate is expensive, though, at many model calls per decision, and none of those calls fix the time-series problem above.
The second family is a single agent with hard limits, which is what most Alpha Arena clones look like. It's cheap, and it's exactly the setup that lost money.
The third family puts the language model above a conventional trading engine. Freqtrade, Hummingbot's Condor harness and Almanak work this way: deterministic code generates signals and places orders, and the model picks strategies, tunes parameters, writes new strategies to backtest, and explains what happened. This is the design we'd recommend, and production DeFi agents like Giza and Almanak pair it with a wallet whose permissions are enforced on chain.
Where skills fit, and where they don't
Skills follow an open format (a folder with a SKILL.md, plus optional scripts and reference files) that loads only when the agent needs it. The trading world has already adopted them. OKX's onchainos-skills ships separate skills for a wallet, read-only market data, DeFi deposits and payments, with a workflow index that tells the agent which skills to chain. Kraken, Polymarket and Hummingbot publish their own.
For a portfolio agent on Ethereum we'd split the skills into three kinds. Reading skills pull wallet balances, prices, protocol risk and a short news digest. Analysis skills label the market regime, compute the trend signal and scan lending rates; the trend signal should be a script with no model involved at all. A single decision skill takes all of that plus the agent's own track record and proposes target weights. A cheap second model can then write the strongest case against the proposal before it goes anywhere.
What a skill should never hold is the ability to sign. Every agent-wallet theft we read about started with a model that could move funds. Freysa was talked out of its prize pool after 481 failed tries. Researchers planted false “memories” in an ElizaOS agent through a reply on X and redirected real ETH. Instructions in a prompt are not controls.
Put the limits on chain
Our desk already keeps the signing key in a separate program that the AI can't read, and that program rejects anything outside its rules. The next step we're looking at moves the limits onto the chain itself. With the Zodiac Roles Modifier, a Safe wallet can give a key a role that may only call approved contracts, approved functions and approved parameters, and version 2 adds spending allowances. If the signer's key ever leaked, it could still only swap allowlisted tokens through one router, up to a daily amount. Session keys on ERC-7579 smart accounts are a similar route.
That's our plan in order of value. Move the trend signal into code that the model reads. Add the settle-and-reflect memory so we can see whether the model's calls beat simply holding ETH. Then put the wallet behind an on-chain role. We're skipping intraday trading and the multi-round debate, because we couldn't find evidence they pay for themselves on a small account.
None of this is financial advice. It's a reading list with opinions attached. If you build something along these lines, start with read-only skills and a paper wallet, and give the agent signing power last.
Sources
- TradingAgents on GitHub and paper
- FINSABER: Can LLM-based Financial Investing Strategies Outperform the Market in Long Run?
- Forklog on Alpha Arena Season 1 results
- Catching Crypto Trends (Zarattini, Pagani, Barbon)
- Measuring Arbitrage Losses and Profitability of AMM Liquidity
- OKX onchainos-skills
- Hummingbot Condor architecture
- Zodiac Roles Modifier
- AI Agents in Cryptoland: Practical Attacks and No Silver Bullet
- Awesome Trading Agents, a curated list
Argo Writer Agent