1980s: Multi-agent systems

Distributed AI and blackboard systems at MIT, CMU and Imperial established the vocabulary we still use: agents, goals, beliefs, intentions. The BDI architecture from Rao & Georgeff in the early 1990s gave us a formal frame for agent reasoning.

Late 1990s: Automated negotiation

Jennings, Faratin and others published the foundational work on bilateral negotiation, utility functions and argumentation-based negotiation. The ideas sat dormant for a decade because deployment required infrastructure that did not yet exist.

2000s: Service-oriented architecture

SOAP, WSDL, UDDI — the first serious attempt at a discoverable, composable service layer. It collapsed under its own complexity but gave us many of the right questions.

2010s: Smart contracts

Ethereum and successors made programmable escrow real. Most agent-commerce use cases did not need a blockchain, but the notion of a binding, auditable exchange baked into the infrastructure became normal.

2020s: Verifiable credentials

The W3C Verifiable Credentials spec, DIDs, and selective- disclosure primitives gave us a practical way to carry scoped consent across organisational boundaries. This is the plumbing for agent consent tokens.

2023–2024: LLM tool use and MCP

OpenAI’s function-calling, LangChain’s tool abstraction and Anthropic’s Model Context Protocol finally gave us a general-purpose way for a language model to act on the world. Once agents could act, the missing piece became a standard transactional layer.

Now: the rendezvous

The five threads converge into something like GeraNexus: BDI goals, negotiation semantics, escrow primitives, verifiable credentials and LLM-driven agents. None of them alone is enough. Together they describe a system that can transact.

A Short History of Research That Led to Agent Commerce