AI Pulse · Under the Radar

Two projects landed this week that show multi-agent trading setups moving from experiment to pattern. Multi Agent Fund split a single crypto wallet across five frontier models, GPT-5.1, Claude, Gemini, Grok, and DeepSeek, giving each one its own desk and a slice of capital on Robinhood Crypto. The code sits in a public GitHub repo. Separately, a YC company called Prodigy Research says its models score in the top 10% of Jane Street traders on internal evaluations. Both announcements dropped within 48 hours of each other.

This is not about one firm trying something weird. It’s about a structure starting to repeat. The Multi Agent Fund approach treats models like portfolio managers, each with its own style and risk tolerance. The idea is that disagreement between agents produces better decisions than any single model acting alone. Prodigy’s claim, if it holds, suggests the models are not just trading on momentum or obvious signals but handling the kind of nuanced, multi-variable problems that separate decent traders from good ones.

The practical question for operators is whether this kind of setup makes sense outside finance. If you run procurement, inventory planning, or pricing decisions across regions, you already juggle conflicting signals and tradeoffs. A multi-agent structure could let you test different strategies in parallel without committing to one model’s blind spots. The Omni Command Centre we build for clients handles exactly this kind of orchestration, routing tasks to the right model and letting you compare outputs before acting.

The evaluations here are internal, so treat the performance claims carefully. But the pattern is real. When two independent teams build the same architecture in the same week, it usually means the idea has legs.