Skip to content
Rush Commerce
AI & Automation2 min read

BNY and Microsoft AI agents negotiate a Treasury bill trade

BNY and Microsoft had AI agents negotiate a U.S. Treasury bill trade inside price limits, with humans approving. The control pattern fits any business.

Two AI agents haggled over a U.S. Treasury bill, and neither one was allowed to close the deal. On November 17, BNY announced that it and Microsoft completed what they call the first agentic negotiation of a U.S. Treasury trade. The agents traded offers. Human traders approved the price. Existing systems booked it. You will never trade T-bills with an agent, but the way this AI agent negotiation was fenced in is a pattern you can copy.

What actually happened

From BNY's press release:

  • The setup: Microsoft Treasury gave its agent trading instructions. The agent connected through a secure gateway to BNY's Agentic Negotiation Protocol (ANP), a patent-pending control layer that lets authorized agents from different companies negotiate.
  • Checks before talk: BNY's controls validated the counterparty, security, size, settlement terms and available pricing. Only then did the agents negotiate, within limits based on approved pricing inputs.
  • Agents cannot execute. Both agents submitted terms separately. ANP checked they matched, then sent them to human traders at both firms for approval.
  • Full record: ANP logs who took part, every message, each control check and each human decision, so the negotiation can be reconstructed.
  • Status: a pilot. BNY plans more testing beyond two-party deals, possibly repo and other contract negotiations. No size, price or model was disclosed.

Microsoft's Ignite 2026 post mentions the trade as part of its BNY work but adds no detail.

Why it matters for your business

Agent-to-agent deals are coming to boring places first: supplier quotes, freight rates, reorders. When your purchasing agent talks to a vendor's sales agent, you need the same four controls BNY used.

  1. Validate before negotiating. Right vendor, right SKU, right quantity, checked by code, not by the model.
  2. Hard price limits from approved inputs. The agent cannot talk itself past a number it was never given.
  3. Separate agreement from execution. The agent proposes. A person or a rules engine you control places the order.
  4. Log the whole conversation. If a deal goes wrong, you need to replay it.

We made the same point when Stripe let agents pay invoices: the approval step is the product. A regulated bank just built its pilot around it.

Key takeaways

  • BNY and Microsoft agents negotiated a U.S. Treasury bill trade in a pilot announced November 17
  • BNY's Agentic Negotiation Protocol checked terms first and held agents to approved price limits
  • Agents could not execute; human traders approved and existing systems booked the trade
  • No trade size, price or AI model was disclosed

Thinking about letting an agent handle quotes or reorders? We build purchasing agents with validation, hard limits, human approval and full logs, on systems you own. See how we build agents or run the numbers on automating your buying.

Sources: BNY press release, Microsoft Ignite 2026.

  • #ai-agents
  • #bny
  • #microsoft
  • #agent-negotiation
  • #human-in-the-loop
TR

Tommy Rush — Founder, Rush Commerce

Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More

Get The Rush Report weekly — one email, zero fluff.