Things Have History
The language between agents

software-architecture

The language between agents

Listen · 4:01

At Google Cloud Next on April 9, 2025, a demo ran on the main stage: a hiring workflow executed by three agents in sequence. One sourced candidates from a job board. Another reviewed résumés against a rubric. A third dispatched interview invitations. None of them were built by the same company. The workflow completed end-to-end because, for the first time, each agent was following the same protocol — Agent2Agent, which Google had announced that morning.

The problem A2A solved is a close cousin of the one MCP had solved five months earlier. MCP addressed the M×N problem for tools: instead of every model learning every API, define one standard connector. A2A addressed the same problem one level up — not between agents and tools, but between agents themselves. An HR agent from Workday and a scheduling agent from ServiceNow had no shared vocabulary. Before A2A, wiring them together meant custom glue code written by a team that belonged to neither company.

The protocol is built on the most unglamorous possible stack: HTTP, Server-Sent Events, and JSON-RPC 2.0. That was deliberate. Google’s engineers, led by principal engineer Todd Segal and product manager Michael Vakoc, chose existing web standards throughout — the same instinct that had made REST durable and MCP adoptable. The three concepts A2A adds are compact. Agent Cards are JSON documents served at a well-known URL, advertising what an agent can do and how to reach it — roughly the role that WSDL played for SOAP services in the early 2000s, except readable. Tasks are structured work items with a defined lifecycle — submitted, executing, completed — that travel between a calling agent and a remote one. Artifacts are the outputs that come back.

The launch came with fifty partners already signed: Atlassian, SAP, Salesforce, ServiceNow, LangChain, Cohere, PayPal, and a roster of consulting firms from Accenture to Wipro. Harrison Chase, the founder of LangChain and one of the engineers who had spent two years building multi-agent orchestration without a standard transport, described the moment with characteristic understatement: “agents interacting with other agents is the very near future.” It already was.

Two subsequent events settled the question of permanence. In June 2025, Google contributed A2A to the Linux Foundation — the same move Anthropic had made with MCP six months prior — placing both protocols under neutral governance. In August 2025, IBM’s competing specification, the Agent Communication Protocol, merged into A2A rather than persist as a separate standard. Protocol consolidation in computing usually takes years; this one took four months.

By early 2026, over 150 organizations had committed to A2A, including Microsoft, AWS, and every major enterprise-software vendor. The question “how do our two agents communicate?” had moved from an integration project to a configuration choice — the same quiet disappearance that TCP/IP made for routing, and REST made for APIs.

Every layer of the stack adds a new M×N problem directly above it. Solve it, and the stack grows a layer taller. The protocols keep arriving.

Sources

Spot a mistake?

Wrong date, broken citation, a fact that doesn't hold? Tell us. It lands in an inbox a human reads and the post can be pulled or corrected.