Microsoft and BNY Mellon Complete First AI Agent-Driven U.S. Treasury Trade
nashnova research
Microsoft and BNY Mellon (纽约梅隆银行) let two AI agents autonomously negotiate a U.S. Treasury trade — a claimed global first — with humans approving only the final step, marking AI's entry into the "high-touch" dealing space long reserved for people.
What did this trade actually involve?
Each side deployed an AI agent that negotiated price and terms on its own, within boundaries set by its human principals. Humans stepped in only to authorize the final execution.
This means → AI's role shifted from "execute a pre-set rule" to "bargain within a framework on my behalf" — the line between algorithmic execution and human negotiation was crossed for the first time.
In plain terms = previously AI could only place orders at prices you specified; now it can haggle with the counterparty, as long as it stays inside your limits.
How did the negotiation work?
BNY Mellon built an "Agentic Negotiation Protocol" — a control layer defining who each agent can talk to, what it can discuss, and where the boundaries lie, plus a bilateral confirmation mechanism to verify both sides read the terms the same way.
Bloomberg reviewed the negotiation logs and noted the two agents used "unusually polite" language — a rarity on Wall Street trading desks.
This reflects an early proof-of-concept stage: the agents have not yet replicated human traders' communication style — they behave more like two very courteous newcomers following a script.
Why Treasuries and not something more complex?
U.S. Treasuries are among the world's most liquid, most standardized assets — the cost of a mistake is relatively contained.
This means → this was a deliberate low-risk drill — prove the concept on the simplest instrument first, then expand. BNY Mellon has already named its next targets: repos, FX payments, collateral terms, and contract negotiation.
In plain terms = you test a self-driving car in a parking lot before taking it on the highway.
What does this mean for market participants?
BNY Mellon's global markets head Laide Majiyagbe said AI agents could lower the barrier to entry in markets where staffing costs previously kept smaller players out.
In plain terms = "high-touch" dealing used to require a full trading desk; if agents mature, smaller institutions could join the game.
He cautioned, however, that the real test is whether the system can scale across asset classes and perform in both calm and volatile markets — this is still the starting line.
What are the key remaining hurdles?
This trade is an early experiment — it has not yet been stress-tested under real market volatility.
How regulators define the liability holder when an AI agent errs — whose fault is it? — will be the central legal question for commercialization.
This means → the technology works in a controlled setting, but between "can do" and "will do at scale" sit two gates: a regulatory framework and extreme-market resilience.
市场有风险,内容仅供研究参考,不构成投资建议。
