The machines were supposed to eat Wall Street by now — instead they're getting schooled in basic market timing.
The Summary
- Anthropic's latest AI agents aimed at Wall Street trading are failing public experiments, showing LLMs aren't ready to replace human traders
- Fund managers remain safe in their jobs as AI struggles with the nuanced decision-making that separates winning trades from losses
- The gap between AI hype and actual trading performance is wider than anyone in Silicon Valley wants to admit
The Signal
Wall Street was supposed to be low-hanging fruit for AI. Pattern recognition, speed, emotionless execution — all the things machines excel at. Yet public experiments testing AI trading bots show they're mostly losing. Not just underperforming. Losing.
Anthropic targeted finance specifically with their latest agent release. They positioned it as Wall Street-ready. The implicit promise: these models could handle the rapid-fire decisions, the risk assessment, the market-feel that separates profitable funds from ones that blow up. The tests are proving otherwise.
"Your fund manager is safe in their job, for now."
The problem isn't compute or speed. AI can process market data faster than any human. The problem is judgment under uncertainty. Trading isn't chess. There's no complete information, no clear win state, and the rules change mid-game when the Fed speaks or a bank collapses. LLMs trained on historical patterns freeze when the pattern breaks.
Consider what trading actually requires:
- Reading between the lines of Fed statements
- Sensing when market structure itself is shifting
- Knowing when to override the model because something feels off
- Managing position size when volatility spikes unexpectedly
These aren't features you can train into a language model. They're emergent properties of human experience, pattern-matching across domains, and the willingness to act on incomplete information. AI agents excel at well-defined tasks with clear reward functions. Financial markets are the opposite — poorly defined problems where the reward function changes based on who else is playing.
The timing matters. This isn't 2023 anymore when every AI demo looked magical. We're in the trough of disillusionment where the technology meets actual work requirements. Bloomberg's reporting suggests even Anthropic's most sophisticated agents struggle with basic market timing.
The Implication
Don't confuse "AI can't replace traders in 2026" with "AI will never replace traders." The experiments failing today are building the dataset for tomorrow's models. Every losing trade teaches the system something. The question isn't whether AI eventually handles trading, it's what hybrid model emerges first — human judgment paired with AI execution, or AI strategies with human override switches.
For anyone building in this space, the lesson is clear. The jobs AI takes first won't be the complex judgment calls. They'll be the repetitive execution tasks surrounding those calls. Trade implementation, portfolio rebalancing, routine hedging. The cognitive core of "what to do" still belongs to humans. The mechanical work of "how to do it" is already slipping away.