Recently, I’ve been looking into the AI trading sector and found that the situation is much more complicated than I initially thought.

Starting in early 2026, after the concept of AI agents became popular, AI trading tools exploded in popularity. In just a few months, over a dozen new projects flooded in, from Nansen launching autonomous trading features, to Donut raising $22 million, to MOSS opening a no-code platform. Major exchanges are all rolling out various skills for AI agents. It looks lively, but a closer look reveals that the line between products that truly manage money and those just storytelling is actually quite blurry.

I’ve roughly divided this sector into three layers. The first layer is pure information tools, with AIXBT as a typical example. It’s like an AI version of a trading radar, posting over 2,000 analyses daily on social media, helping you identify promising targets, but not placing trades for you. These products are growing the fastest because they are the furthest from capital security.

The second layer is the core focus of this round—decision-making combined with execution. Minara offers four ways to place orders, from manual to fully automated, supporting short-term, intraday, and swing styles. Donut has become a browser-level operating system, allowing you to trade directly while viewing candlesticks or browsing DEXs. MOSS lets you describe strategies in plain language, turning AI into a trading agent, but interestingly, it first puts the agent into “hell mode,” using 150 days of real historical data for stress testing; only after passing can it go live. Nansen relies on its advantage of over 500 million labeled wallet addresses to monitor on-chain anomalies and execute trades directly.

There’s also Mojo AI innovating in DeFi, supporting natural language commands for swaps, cross-chain transfers, staking, and more. Cod3x runs on perpetual contracts, with a product called Big Tony demonstrating an excess return of 21.7% compared to holding BTC. Milo operates a non-custodial agent on Solana, with the interesting feature that each transaction is accompanied by a “trade journal” explaining the logic, offering good transparency. HyperAgent charges $550 per month, analyzing seven signals simultaneously, with 17 hardcoded safety restrictions, but its user base is small.

The third layer is infrastructure. VergeX’s NoFx is an open-source project that can connect to multiple exchanges, not limited to cryptocurrencies. Almanak is more aggressive, employing 18 specialized AI agents working collaboratively, with over $10 million in funding.

But the risks behind all this are also quite clear. First, systemic risk: many AI agents use the same large models, with highly similar analysis standards. If a certain condition triggers, thousands of AIs might sell simultaneously, causing a collective crash. Some projects are trying to break this, like HyperAgent using dynamic weighting across seven signals, and Almanak employing multi-brain decision-making, but how much these can truly mitigate extreme market conditions remains to be tested.

Second, the proliferation of “pseudo-AI”: many so-called AI trading platforms are still running traditional technical indicator scripts, just wrapped in an AI shell. Users think they’re trading with AI, but they’re actually using old bots with a new package.

Third, AI itself can “talk nonsense”: inventing nonexistent trading pairs, misreading on-chain data, giving outdated judgments during volatility—these can lead directly to real financial losses. Even more dangerous are prompt injection attacks, where hackers embed malicious commands into code or web pages. If AI executes these without discernment, the consequences could be disastrous. That’s why most products still retain manual confirmation steps, but that also causes missed trading opportunities.

Fourth, strategies tend to fail in bear markets. Most models are trained on historical data and may become ineffective in new market conditions. AI works best under the assumption that “history repeats itself,” but markets are excellent at breaking that premise.

So, before being swayed by the story of “AI helping you trade crypto,” it’s best to ask three questions: Is it really AI, or just a shell for old scripts? Who is holding your money? And from just watching the charts to actually managing money, and doing it well—what’s in between is not just code upgrades but a long journey of building trust.