AI Agents Are Starting to Negotiate With Each Other Directly

A capability that was mostly a research demo eighteen months ago — one AI agent negotiating a transaction directly with another AI agent, no human relaying messages in between — is now showing up in real, if still small-scale, commercial deployments. Procurement agents are requesting quotes from supplier agents, receiving counter-offers, and settling on terms within pre-authorized bounds, with a human only reviewing the completed transaction log rather than approving each step. The unlock has been standardized agent-to-agent communication protocols that specify not just message format but negotiation semantics: how an offer is structured, what counts as a counter-offer versus a rejection, and how a deal is finalized in a way both sides' systems can verify without disputing the outcome afterward. Without that shared protocol, agent-to-agent negotiation degenerates quickly into brittle, one-off integrations that break the moment either side changes its internal logic. The current deployments are deliberately narrow: bounded price ranges, pre-vetted counterparties, and hard caps on transaction size, with anything outside those bounds automatically escalating to a human. That's a sensible guardrail given the failure mode — an agent with a subtly miscalibrated reward function negotiating against another agent can drift into bad deals faster than a human would ever notice, since neither side is pausing to sanity-check the exchange the way two human negotiators naturally would. For platforms that already broker transactions between many independent parties — marketplaces, classifieds, B2B sourcing — this is the capability to watch. The infrastructure question isn't whether agents will negotiate on a platform's behalf, but how soon, and how the platform verifies that an AI-negotiated deal was actually fair to both sides.