BNY and Microsoft have completed a pilot in which software agents representing each firm negotiated the terms of a U.S. Treasury bill trade, according to a community report on dev.to republishing Fathom News. BNY described the exercise as an industry-first negotiation of a Treasury bill trade conducted by AI agents. The account is a single attributed source, so the details below reflect what BNY and the report state rather than independently verified results.
The mechanism BNY describes is its Agentic Negotiation Protocol, or ANP. The protocol carried the back-and-forth between the two organizations' agents, with each side exchanging offers and counteroffers inside predefined rules and approved pricing inputs. Before any proposed terms reached a human, ANP checked a central ledger and confirmed that both parties' agents had submitted matching terms.
On the participant side, controls were present too. Specific trading instructions were handed by Microsoft Treasury to its agent, whereas checks by BNY covered the counterparty, the security, the size, the settlement terms and available pricing — and this happened before negotiation of prices within authorized limits could occur. ANP is framed in the report as a control layer meant to let authorized agents from different organizations negotiate securely and transparently, verifying participants and giving them a secure connection.
What the agents could not do is the most consequential detail for anyone watching automation in financial workflows. Proposed terms were reviewed and approved by human traders, and execution and booking of the trade went through the firms' existing internal systems and controls. For either firm, the agents could not execute the trade. Negotiation and term-matching, not autonomous execution, were therefore what the pilot tested.
The process additionally generated an audit trail. Per the report, a record is maintained covering participants, the messages they exchanged, control checks and human decisions, enabling a negotiation to be reviewed and reconstructed afterward. BNY's stated aim is to make the conventional message-driven workflow more efficient without disturbing the risk and transaction-processing safeguards already in place.
BNY's scale supplies context for the pilot. The report states that more than 90% of Fortune 100 companies, along with nearly all of the top 100 banks globally, are served by BNY; the institution reported custody and/or administration assets of $63.4 trillion and managed assets of $2.2 trillion, both measured at the end of September. Those figures describe the institution rather than the pilot's scope, which the report calls an initial bilateral use case.
Further testing of ANP is planned by BNY, which intends to expand beyond that first bilateral case and to evaluate additional negotiated workflows. Repo and other contractual negotiations could be among future applications over time. No dates for those next tests were given by BNY, the report notes, so the timeline for any broader rollout remains unspecified.
The report does not disclose the size of the Treasury bill trade, the number of negotiations run, error rates, or how the agents' performance compared with the traditional message-based process. It also does not describe the underlying models or how the approved pricing inputs were generated. Those gaps matter because the claim of an industry first rests on BNY's own characterization.
For developers and freelancers who build or integrate agent systems, the pilot's design is the transferable part. The agents operated inside a bounded envelope: predefined rules, approved pricing inputs, participant verification and a matching check before escalation. Authority to commit stayed with humans, and execution stayed with existing systems. That is a pattern of constrained autonomy rather than open-ended agent behavior.
The audit requirement is equally practical. Recording participants, messages, control checks and human decisions turns an agent negotiation into something reviewable after the fact. Teams building similar workflows would need logging and reconstruction designed in from the start, not bolted on, if the goal is to satisfy risk and transaction-processing controls that already exist.
There is also a tradeoff worth naming. The report says the approach is intended to streamline a message-based process, but the pilot retained human review and existing booking systems, and it added verification and matching steps. Whether the protocol reduces elapsed time or manual effort is not established in the supplied material; the report describes intent and design, not measured outcomes.
The expansion plans point to where the harder problems sit. Repo and other contractual negotiations involve different terms and documentation than a Treasury bill trade, and the report gives no dates or scope for those tests. Treat the broader applicability of ANP as a stated intention rather than a demonstrated capability.
The honest conclusion is narrow. BNY and Microsoft ran a bilateral pilot in which agents negotiated within rules and humans approved the terms, with execution handled by existing systems. That is a concrete, attributable event, but the evidence supplied does not include independent verification, performance data or a published technical specification of the protocol.
For this audience, the useful takeaway is editorial interpretation: the pilot illustrates a governance-first approach to multi-party agents, where identity checks, rule-bound negotiation, term matching and human sign-off are treated as prerequisites rather than afterthoughts. Developers evaluating agent frameworks for regulated or high-stakes workflows can use that checklist, while recognizing that the source here is a single attributed account of one pilot.