deniz.in

Markets

Weather

Loading weather

· via dev.to (home feed)

BNY and Microsoft complete first AI agent negotiation of a US Treasury bill trade

BNY and Microsoft ran a pilot in which AI agents negotiated a US Treasury bill trade using BNY's Agentic Negotiation Protocol, with human traders approving the terms and existing systems handling execution.

BNY and Microsoft complete first AI agent negotiation of a US Treasury bill trade

What happened

BNY and Microsoft have completed a pilot in which AI agents from the two firms negotiated a U.S. Treasury bill trade, an arrangement BNY characterizes as an industry first for agent-to-agent dealmaking in this market. According to Fathom News, the exercise was built to test whether governed AI agents can handle the offer-and-counteroffer stage of a trading workflow without taking humans out of the loop.

The negotiation ran on BNY's Agentic Negotiation Protocol, known as ANP, which the bank designed as a control layer that lets authorized agents from different organizations negotiate securely and transparently.

How the negotiation worked

Microsoft Treasury supplied its agent with specific trading instructions. On the BNY side, before any price talk could begin, controls validated the counterparty, the security being traded, the trade size, the settlement terms and the available pricing inputs.

Once those checks passed, the two agents exchanged offers and counteroffers within predefined rules and authorized pricing limits. Participant-side controls combined those rules with AI safety checks tied to approved pricing sources. ANP verified the participants, established secure connections between them, and consulted a central ledger before proposed terms were shown to traders. As a final gate, the protocol confirmed that both parties' agents had submitted matching terms before requesting human approval.

Humans kept final authority

The agents could not execute the trade on either side. Human traders reviewed and approved the proposed terms, and execution and booking went through the firms' existing internal systems and controls — the same path a manually negotiated trade would follow.

The process also maintained a full record of who took part, the messages exchanged, the control checks performed and the human decisions made, so the entire negotiation can be reviewed and reconstructed after the fact. BNY said the aim is to streamline the traditional message-based negotiation process while leaving existing risk management and transaction-processing controls intact.

Why it matters

The pilot is a concrete example of agentic AI operating inside one of the most tightly controlled environments in finance. Rather than asking whether agents can replace traders, BNY's design splits the workflow: agents do the repetitive bargaining within hard limits, while humans retain approval and execution. The audit trail — participants, messages, checks and decisions — addresses the reconstructability requirements that regulators and risk teams expect from electronic trading.

Scale makes the test notable. According to Fathom News, BNY serves more than 90% of Fortune 100 companies and nearly all of the top 100 banks globally, and as of September 30 it oversaw $63.4 trillion in assets under custody and/or administration along with $2.2 trillion in assets under management. A negotiation protocol proven at that scale could shorten the dealmaking cycle in markets where price discovery still depends heavily on chat messages and back-and-forth human communication, and it offers a template other institutions could adopt for inter-organizational agent workflows.

It also reflects a broader shift in how financial firms deploy AI: not as autonomous decision-makers, but as governed participants whose authority is bounded by validation checks, matched-term verification and mandatory human sign-off.

What comes next

BNY plans to continue testing ANP beyond the initial bilateral, two-party setup and to evaluate additional negotiated workflows. Future applications the bank is considering include repurchase agreements and other forms of contractual negotiation. No timeline was given for those next tests.

  • #ai-agents
  • #microsoft
  • #bny
  • #finance
  • #automation

Related posts