Can AI agents actually make money trading?

A short, skeptical read on whether LLM-driven agents are delivering anything close to the "near-guaranteed" returns the HFT comparison implies — and where the genuine edges, when they exist, actually live.

TL;DR

LLM agents are a real category in 2026. Academic frameworks like TradingAgents and StockBench show they can squeeze a small return out of equity trading. But:

The best honest benchmark in the literature, StockBench, finds the top LLM agent (Kimi-K2) returned 1.9% over the test window vs. 0.4% for a buy-and-hold baseline — a real edge but tiny (StockBench, 2025).
All LLM agents failed to beat buy-and-hold in bear markets. They mostly outperform when markets rise.
The "near-guaranteed" comparison to HFT arbitrage breaks down the moment you ask what makes HFT near-guaranteed: microsecond co-location, audited regulatory privileges, structural market-making relationships, billions in capital, and physics-grade infrastructure. A retail person running an LLM agent has none of that.
The one place where agentic returns look genuinely outsized is prediction markets — Polymarket-style binary outcomes with limited liquidity — where individual trades can return 50–400% and 30%+ of wallets are now AI-driven.
For regular stocks and crypto: assume 5–15% annualised at best with a well-configured bot, occasional drawdowns, and that any "guaranteed 70%/year" pitch is either a backtest or a scam.

What HFT firms actually do — and why "guaranteed" is misleading

The framing that HFT arbitrage is "almost always winning" is a half-truth that comes from the way Michael Lewis told the story in Flash Boys. The reality in 2026 is more nuanced.

Pure latency arbitrage — buy on exchange A at $100.00, sell on exchange B at $100.02 in the same microsecond — was a real business model in the 2010s and is now largely played out. Markets are too well-connected, regulators have tightened the rules, and the firms that still do it operate at speeds that require microwave towers between Chicago and New Jersey, co-located servers inside the exchange, and FPGAs that respond in nanoseconds. Costs of entry are in the tens of millions.

What the big firms — Citadel Securities, Jane Street, Hudson River Trading, Jump Trading — actually do in 2026 is a mix of:

Market-making: continuously quote both bid and ask on thousands of instruments, capturing the spread, eating the inventory risk. Profitable on average, lossy on any given day if you mishedge. Citadel Securities and HRT are dominantly market-makers (Bloomberg on HRT, Dec 2025).
Statistical arbitrage: identify a statistical relationship between assets, trade when it temporarily diverges, profit when it reverts. Not risk-free; the relationship can permanently break.
ETF arbitrage: Jane Street's specialty — capture the gap between an ETF and the basket of underlying securities. Holding periods of minutes to hours, not microseconds.
Hedge-fund-style proprietary trading: HRT now reports an average 5-minute hold time and keeps 25% of positions overnight (Bloomberg, 2025). That's quant trading, not arbitrage.

What makes these firms reliably profitable is structural — co-location, capital, exchange relationships, regulatory licensing — not magic algorithms. None of those are accessible to a developer running an LLM agent on a VPS.

What LLM agent trading actually looks like in 2026

Strip away the marketing and there are three actually-existing patterns:

1. Multi-agent LLM frameworks for equity research and trading. The most prominent open-source example is TradingAgents (Tauric Research, GitHub) — a framework where multiple LLM-driven agents play specialised roles: a fundamental analyst, a sentiment analyst from news/social, a technical analyst, "Bull" and "Bear" researchers, a risk manager, and a trader who synthesises the verdict (TradingAgents project page). It works in the sense that it produces trade decisions and backtests well on cherry-picked tickers (AAPL, GOOGL, AMZN). Whether it makes money for someone running it live is a separate question with no audited answer.

2. Academic multi-agent systems with bold headline numbers. The 2025 HedgeAgents paper claims a 70% annualised return / 400% over three years with a balanced-aware multi-agent system (arXiv:2502.13165). A Review of LLMs for Stock Price Forecasting paper accepted at IEEE in 2026 catalogues a dozen similar systems (arXiv:2605.05211). The pattern across these papers: large reported returns, in-sample or short-window backtests, no live-trading attestation, and known issues with data leakage (the LLM has read the news that "happened" after the trade date) and survivorship bias (the tickers it traded on happen to be the ones that went up).

3. Honest benchmarks that test out-of-sample. The most useful single source is StockBench (2025, arXiv:2510.02209), which evaluates LLM agents in realistic real-world equity environments with no look-ahead. The verdict is sober: the top performer (Kimi-K2) returned 1.9% in the test window, the average buy-and-hold baseline returned 0.4%, the margin is positive but tiny, and in bear markets every tested model failed to beat the passive baseline (StockBench). One useful finding: LLM agents do limit drawdowns better than buy-and-hold (–11% to –14% vs the baseline's –15%), so they trade some upside for downside protection. That's a real characteristic but not "near-guaranteed wins."

The honest read: LLM agents in 2026 are better than chance, marginally better than buy-and-hold in normal markets, and worse than buy-and-hold in bear markets. They aren't fake, but they aren't the HFT comparison either.

The one area where agent returns look genuinely high: prediction markets

The exception that proves the rule. On Polymarket and similar prediction markets, individual events have bounded outcomes, limited liquidity, and a constant influx of news. That's exactly the shape of problem an LLM can engage with profitably: read the latest reporting, update probability estimates, place trades when market prices diverge from the model.

Olas's Polystrat agent executed more than 4,200 trades on Polymarket in a single month and reportedly hit returns as high as 376% on individual trades (CoinDesk, March 2026). Over 30% of wallets on Polymarket are now AI-driven. This is genuinely a category where LLM agents are working, and the reason is structural: prediction markets are slow, inefficient, news-dense, and small enough that no HFT firm bothers.

The catch: prediction-market scale is small. You can't deploy $10 million into a single Polymarket contract — the liquidity isn't there. So the strategy works for tens-of-thousands of dollars at a time, not for fund-scale capital. For a developer playing with $1k–10k, it might be the most interesting niche in this space right now.

The retail "agentic brokerage" pitch — and what to be wary of

Brokers like Public.com are now selling "agentic" trading: you describe your strategy in natural language, AI agents execute it continuously, you stop pushing buttons (Public.com AI Agents). The pitch is genuine; the implementation is real; whether it makes money is the same question as every other retail trading product.

The widely-cited industry statistic is that about 60% of retail algorithmic traders show positive annual returns, but fewer than 1% of all day traders (manual or automated) consistently profit net of all fees (TradingView Hub guide, 2026). Both numbers are likely overstated. The 60% figure is usually from self-reported user data on bot-vendor sites, which is selection-biased toward people who haven't yet quit.

Realistic expectation ranges from the same sources: 5–15% annualised for beginners, 15–25% for experienced operators with proven strategies, in good market conditions. That is fine, but it is the same range you can hit with a low-fee index fund in a good year, with vastly less work and risk. Any "AI-powered bot" advertising 40%+ annualised guaranteed, or showing screenshots of equity curves that go up and to the right with no drawdowns, is selling either a backtest or a fraud.

The deeper problem: where does an LLM's edge actually come from?

The question worth sitting with: if an LLM trained on public text reads public news and produces a trade decision — what edge does that have over the hedge fund analyst who is doing the same thing, with better data, more capital, faster execution, and twenty years of experience?

The honest answer is: not much. The market is competitive precisely because so many smart people are looking at the same public data. Any edge an LLM extracts from public information will be small, decay quickly as others find the same trick, and disappear entirely if the strategy gets crowded.

Edges that survive in retail-accessible markets in 2026:

Speed of synthesis on news-dense, illiquid markets (Polymarket and similar prediction markets — works).
Discipline and consistency in following a pre-defined strategy (the LLM doesn't panic-sell; this is real, but it's an execution edge, not an alpha edge).
Niche markets with thin attention (some altcoin pairs, some emerging-market ETFs, some retail-flow-driven names) — works briefly, until enough agents are doing the same thing.

Edges that don't survive:

"Read Twitter and predict the S&P." Every fund has been doing this since 2015. Sentiment is priced in.
"Backtest on historical data with LLM-generated trades." Data leakage from the LLM's training set destroys this category. The model has read what happened next.
"Replicate HFT arbitrage." See above — that game requires hardware and licensing you don't have.

Bottom line

Are AI agents making money in trading? Yes — modestly, in narrow domains, in specific market conditions. Are they delivering anything close to HFT-style near-guaranteed returns? No, and the comparison is misleading: HFT is profitable because of structural advantages, not algorithm magic. The honest StockBench number — 1.9% vs 0.4% baseline, fails in bear markets — is what to anchor on for stocks.

Where to look if you want to play:

Prediction markets (Polymarket via something like Olas/Polystrat) — the one category where LLM agent returns look genuinely interesting, because the markets are too small and slow for the big firms to bother.
A serious, defensively backtested rules-based bot running on your own infrastructure (3Commas, WunderTrading, or custom Python on CCXT) targeting 5–15% annualised. Treat any returns above that range as suspect.
TradingAgents on GitHub as an excellent educational project — read it, run the backtests, learn how multi-agent prompting fits with finance. Don't put real money behind it.

What to avoid:

Any platform advertising guaranteed returns or 40%+ annualised audited live performance with no drawdowns.
Pump-and-dump signal services with an "AI" coat of paint.
Trying to replicate HFT strategies on a retail broker. The latency math alone makes it impossible.

The closest honest summary in 2026: an LLM-driven trading agent is a thoughtful intern with no money of its own — useful for sifting information, helpful for execution discipline, occasionally insightful, never the source of structural alpha.

Sources

Numbers verified May 2026. Trading-bot vendor pages are aggressive marketing; treat any self-reported return figure with suspicion until you see audited live performance.