Why No Expert Would Recommend AI Trading Bots

No major AI company has endorsed crypto trading bots. No frontier lab is training models for it. Yet a growing number of traders are using Anthropic’s Claude to build automated Polymarket bots, claiming profits in the millions. Viral threads suggest anyone can do it.

But the loudest winners are using strategies any quant fund could replicate overnight.

Three Assumptions, Zero Guarantees

The narrative rests on three assumptions. Big tech will eventually build purpose-made trading models. Individual traders can sustain an edge against institutions. Autonomous AI agents can reliably make money in open markets.

Haseeb Qureshi, managing partner at Dragonfly Capital, disagrees on all three counts. In the Bankless interview, he pointed to liability risk, market structure, and the commoditized nature of AI. Together, these forces make this gold rush far less promising than it appears.

The Liability Trap

Qureshi says that building AI for blockchain tasks is technically trivial. An EVM simulator can test looped lending or token swaps with ease. The models are capable. They just haven’t been pointed at crypto.

The reason is institutional, not technical. First, crypto carries reputational baggage that AI labs want no part of. “Crypto’s kind of cringe,” Qureshi said.

But the real barrier is liability. Imagine Claude botches a leveraged trade and wipes out $2 million. Or sends $10,000 to a burner address by mistake. No disclaimer would be loud enough to prevent the backlash.

“It will 100% happen,” Qureshi said. “Anybody who has a bad experience, it’s going to go super viral.”

He compared managing a user’s crypto wallet to injecting unregulated Chinese peptides. The downside dwarfs any revenue upside. Coding advice gone wrong is embarrassing. A drained wallet is a lawsuit.

Anthropic has already published research on AI and blockchain. Its SCONE-bench study tested how well frontier models exploit smart contract vulnerabilities. But this is cybersecurity research, not a product roadmap.

The inflection point will come from competition. When one lab decides crypto volume is too strategic to cede to rivals, training will begin. Until then, silence.

The Jane Street Problem

Even without big tech, the trading narrative faces a structural wall. Any strategy built on a publicly available model is, by definition, available to everyone — including institutional quant firms.

Qureshi’s point is simple. If a basic Claude bot can find profitable trades on Polymarket, Jane Street can run 5,000 of them simultaneously. The firm has faster infrastructure and deeper capital. It can scale any profitable edge to zero before a retail trader even logs in. “If it’s in the raw model, Jane Street is doing it right now,” he said.

The only way a retail bot wins is with novel signals absent from the base model. A Claude instance pointed at an API is not that.

Why ‘Go Make Money’ Doesn’t Work

Qureshi extended the argument beyond trading to the broader fantasy of autonomous AI agents earning income on their own.

The first option is getting hired — having the AI agent sell its labor. But this is economically impossible. Millions of identical Claude instances exist. None has a unique skill or location advantage. Hiring an AI agent is just buying Anthropic compute with extra steps. No rational buyer would pay above Anthropic’s own API price for the same output.

The second option is starting a business. This sounds more promising, but Qureshi argued it fails for a subtler reason. Every AI agent draws ideas from the same pool of training data. The result is that they all converge on the same generic plans. Ask ten Claude instances for a startup idea, and you get ten variations of the same pitch.

Real entrepreneurship, Qureshi said, requires what Peter Thiel calls “earned secrets.” These are insights born from specific experiences in specific places at specific times. Bankless built its brand because its founders had a unique mix of crypto expertise, storytelling, and community instinct. They had it at exactly the right moment. A freshly spun-up Claude has no life experience to draw from. It has no earned secrets.

This leads to an uncomfortable conclusion. AI agents cannot win at trading. They cannot get hired. They cannot generate original business ideas. So, where is their genuine advantage over humans? Qureshi’s answer was deliberately provocative: crime. This is not a future Qureshi welcomes. It is where the logic leads when you remove every institutional guardrail.

What This Means

The traders building Polymarket bots are real. Some profits may be real, too — for now. But institutional quant firms will arbitrage away any alpha in the base model. Big tech will not train on crypto until forced by competition. And the autonomous agent economy may find its first viable model beyond law enforcement’s reach.

For the average trader reading headlines about AI bots minting millions, the takeaway is implicit. The house always wins. In AI trading, the house runs 5,000 bots with sub-millisecond latency.