AGTBRIEF
An arXiv preprint proposes multi agent AI for cross border clinical stage biotech, where cash flow models don't apply. No AI is evaluated; the cited 127.17% return is the human author's prior record.
A pre-revenue clinical-stage biotech can roughly halve or double on a single trial readout or FDA decision, so the cash-flow inputs an analyst would normally use do not exist. A new arXiv preprint by Yuhan Fang, submitted 10 August 2026, treats the missing inputs as a design problem for multi-agent AI: a team of LLM agents working in specialist roles. The paper proposes a three-layer architecture to handle them.
The framework's first layer converts qualitative scientific judgment into defensible price ranges for pre-revenue assets; the second reconciles listings on different international venues at the same moment; the third arbitrates bullish scientific conviction against cautious regulatory constraints. The full text introduces a "Glocal" practice spanning founder capability, cross-border regulatory velocity, and vehicle liquidity.
The caveat sits at the top: the paper presents an architecture, not a system, and no AI implementation is evaluated. The headline 127.17% return over sixteen months against a 50.67% benchmark is the author's own manual track record as sole portfolio manager of China's first dedicated cross-border biotech fund (摩根中国生物医药混合QDII A), publicly registered in China under fund code 001984 (fund manager page): provenance for the cited performance, not evidence that any AI ran the book.
What remains open: an independent benchmark of the architecture, an audited confirmation of the human track record, and any sign a fund, bank, or lab is building it.