The AI agents showing up this year

A chat agent

A chat agent reads prices, news, or both and picks the trade in the moment. Research systems in that family include FinMem, FinCon, TradingAgents, and QuantAgent. Their historical tests often look excellent. A leakage study compared the same systems before and after the underlying model's knowledge cutoff. Returns fell about 50–72 percent. Sharpe ratios fell about 51–62 percent. The model was recognizing history it had already read.

Live benchmarks exist because a backtest on old prices can smuggle the future in through memory. A later test that hid company names found the headline profit was mostly the market and the style the agent happened to hold. Stock-picking skill landed near zero, and for most of the models it was negative. A strong agent backtest is a prototype. A live account asks for a record the prototype has not produced. The measurements are in Profit Mirage, the case for a live benchmark is in DeepFund, and the deployment warning is in The Alpha Illusion. A memory-controlled stock test is written up separately as From Knowing to Doing.

A score, or a bot you still design

A signals desk publishes a rating or a menu of bots and calls the engine AI. Trade Ideas is the retail stock screener with the longest public simulation, under the name Holly. A simulation still assumes a fill. The account pays the spread. Tickeron is a menu of separate bots. Read that kind of menu one bot at a time. A backtest on a marketing page is a hypothesis. The record that matters is a forward log of the bot you would actually fund.

An automation platform is older, and clearer about being a machine. 3Commas runs grid bots and repeated-buy bots on crypto exchanges, and has added QuantPilot, an agent that proposes strategies and backtests them. Composer lets you assemble if-then rules, mostly for US funds, and sends the orders. These can sit on an account and act. Their quality is the rule you left running. A repeated-buy bot follows a falling market down when that is what it was told to do. A backtest that fills at the daily close and ignores the spread prints an easier path than the account will see.

What this platform sends

A chat agent is as good as its forward record, and the flattering tests overlap what the model already read. A signals product is as good as the log of the specific bot you would fund. An automation platform is as good as the rule you left on, including on the days that rule keeps buying.

This platform publishes the grade, the card, and the drawdown. The scanner covers the same 221 stocks and 35 crypto names. A setup earns an A or an A+ or the card is never sent. The card names the entry, the stop, and the targets before the order. The bot places that card on the account you own, inside the position count you chose. The grade stays the grade after a losing week.

In the published simulation from 1 November 2023 through 27 September 2026, this platform's investor simulation had a maximum drawdown of −6.6%. Buy-and-hold SPY drew down −18.8% in that window. The figure is a simulation. The research page shows the window, the return, and the comparison together. Keys on a connected exchange are created without withdrawal permission. Demo comes first. A kill switch stops new orders. Those controls bound an operational mistake. The market can still move through a stop.