The experiment started simple enough: a physics researcher—no computer science degree, no finance training—hooked Claude up to a US brokerage account via API and let it trade autonomously for one week. The setup was deliberately minimal. No fancy retraining pipelines, no proprietary indicators, just an LLM connected to the market with a basic automated strategy executing in real money.

What Actually Happened

The conversation exploded across Chinese AI developer communities this week, not because of spectacular gains or catastrophic losses, but because of what the experiment revealed about the gap between 'LLM as agent' and 'LLM as actually useful trading tool.' The researcher documented everything: prompt engineering choices, decision logs, execution errors, and the uncomfortable moments where Claude had to reason under uncertainty.

The Core Question Nobody Wants to Answer

Here's what makes this interesting: a plain Python script with if-this-then-that logic could execute the exact same trading strategy. Move averages, threshold-based buy/sell triggers—these don't require natural language understanding or chain-of-thought reasoning. So what exactly does Claude add? The answer isn't obvious, and that's the point.

Where LLMs Might Actually Help

The potential value isn't in executing known strategies; it's in adapting them dynamically. When market conditions shift unexpectedly, a scripted bot follows its rules until someone updates it. An LLM agent could potentially recognize novel patterns, adjust position sizing based on confidence, or synthesize external information into trading decisions without human intervention.

The Reality Check

But let's be real: one week of live trading with no baseline comparison tells us almost nothing about profitability. What it does tell us is that the technical barriers to entry have collapsed. Anyone with API access and basic prompting skills can now deploy AI agents in production financial systems—a capability that required teams of engineers just a few years ago.

Key Takeaways

  • The experiment proves accessibility, not efficacy—connecting an LLM to live markets is now trivial
  • Traditional scripts outperform LLMs on pure execution speed and cost efficiency for known strategies
  • LLMs add value in adaptive reasoning but introduce latency, inconsistency, and unpredictability
  • Regulation and risk management frameworks haven't caught up with AI agents in finance

The Bottom Line

This isn't a win for Claude or a loss—it's a stress test of our assumptions. The researcher didn't need to be an expert because Claude handled the complexity; but that same flexibility is what makes it hard to trust with serious capital. The honest answer to 'what does the LLM add?' might just be: everything you didn't ask for.