AINEWS
The Open Agent Protocol Is the Least Important Layer in the Multi-Agent Stack
Agent2Agent (A2A) is a vendor neutral interoperability standard for multi agent AI.
Agent2Agent (A2A) is a vendor neutral interoperability standard for multi agent AI.
Google's Agent2Agent (A2A) protocol is, on paper, the most polite thing in the AI standards fight. The spec is, on its face, open. It is genuinely cross-language. A Python agent and a Go agent can exchange work through A2A without ever touching a Google service.
That is the layer to watch, because it is also the layer that does not matter.
The contested territory sits one level up and one level down. A2A is the wire. Above the wire is the orchestrator that assembles agents into a pipeline. Below the wire is the model that does the cognitive work and the cloud where both run. The June 22 tutorial on Google's developer blog, titled Build Cross-Language Multi-Agent Team with Google's Agent Development Kit and A2A, ships a Contract Compliance pipeline that uses A2A for the wire, the Agent Development Kit (ADK) for orchestration, and Gemini for the only nondeterministic step in the system. The Go agent handles the deterministic compliance check. There is no architectural reason for that split between languages. It is Google's parts list.
The framing of the agent interop race as a contest between open protocols and proprietary frameworks is the trap. The protocol is the layer that least determines who wins, because the protocol is supposed to be free. Whoever ships the most useful orchestrator, the most capable reference model, and the most frictionless deployment substrate captures the value that the protocol appears to surrender.
Google is positioning itself to ship all three — ADK as the orchestrator, Gemini as the reference model, and Google Cloud as the implied deployment substrate — at a pace it presents as unmatched by consensus standards or rival frameworks. The Google Developers Blog is the tutorial on-ramp. A2A is the open door. Every well-marked path through that door is Google-shaped.
Some observers have noted an analogous pattern in Google's approach to gRPC: the protocol was donated to a foundation, but the supporting tooling, documentation, and default language bindings pointed back to Google Cloud. By the time governance stabilized, the defaults were the standard. A2A is being seeded through the same mechanism: developer blog posts, sample code, and worked examples that all happen to assume ADK and Gemini. The tutorials are free. The architectural choices they normalize are not.
The risk is not that A2A will be hijacked. A2A does not need to be hijacked. It only needs to be surrounded. When the canonical example for a cross-language agent pipeline is Google's example, and the canonical orchestrator is ADK, and the canonical model is Gemini, neutrality at the wire has very little to do with who controls the pipeline at the end of the day.
The same June 22 post makes a quieter case worth reading. The author names real production failure modes in monolithic agents: context degradation past roughly 10 to 15 tools, total blast radius on any unhandled exception, and an untestable entanglement of responsibilities inside one giant prompt. The fix the post proposes is narrow-responsibility agents coordinated by an orchestrator, and the cross-language collaboration pattern is one of three the post promise to teach. That architectural lesson is durable. The lesson survives ADK and survives A2A. What does not survive is the assumption that any team picking up the pattern will reach for anything other than Google's stack by default.
What to watch next. If the next batch of third-party multi-agent tutorials assume ADK and Gemini without comment, the standards race is over before governance has caught up. If a meaningful number ship on a different orchestrator, a different model, and a different cloud, with A2A still as the wire, the open protocol will have done its job. The wire layer is decided. The stack around it is still being written, and the example being written into it is Google's.