This is a submission for the Google Cloud NEXT Writing Challenge
I watched every hour of Google Cloud NEXT '26. And like everyone else, I was dazzled by the Gemini Enterprise Agent Platform rebrand, the 8th-gen TPUs, and the Workspace Studio demos.
But the announcement I keep coming back to — the one I think will quietly matter the most to developers building real things — is something that didn't even get its own keynote slide.
Two protocols. A2A and MCP. And how Google just bet the entire agentic era on both of them.
A Quick Reality Check First
Let me say the quiet part out loud: the "agentic enterprise" pitch Google delivered at NEXT '26 is compelling, but it's also a vision. Agents that autonomously coordinate across your entire org, spinning up sub-agents, handing off tasks, self-healing — we're not fully there yet for most teams
MCP: You Probably Already Know This One
By now, if you're building with LLMs, you've likely heard of Model Context Protocol (MCP) — Anthropic's open standard for connecting agents to tools, APIs, and data sources. It answers the question: "How does my agent talk to the outside world?"
At NEXT '26, Google announced managed MCP servers across Google Cloud services and used its Apigee API management platform to turn any existing API into a discoverable, agent-ready MCP tool. That's a big deal for enterprises that have years of internal APIs sitting around — they don't need to rewrite them. They just need to surface them via MCP.
But here's what I think most people missed: MCP only solves half the problem.
A2A: The Protocol Nobody Is Talking About
MCP connects an agent to tools. But what happens when you need agents to talk to other agents?
That's what the Agent2Agent (A2A) protocol is for. Announced by Google last year and now at v1.2, A2A was quietly handed off to the Linux Foundation at NEXT '26 — a signal that Google is treating this as genuine open infrastructure, not a proprietary lock-in play.
Here's the practical scenario A2A unlocks
Agent A (Gemini, on Google Cloud)
↕ [A2A]
Agent B (Claude, on AWS Bedrock)
↕ [A2A]
Agent C (custom model, on-prem)
These three agents — built on completely different models, by different teams, running on different clouds — can now hand off tasks, share context, and collaborate. Without any of them needing to understand each other's internals.
Google confirmed A2A is already in production at 150 organizations. Native A2A support ships inside Google's Agent Development Kit (ADK), and it's already integrated with LangGraph, CrewAI, LlamaIndex, Semantic Kernel, and AutoGen.
My Honest Take: The Risks
I want to be balanced here, because the dev.to community deserves more than hype.
The governance question is unresolved. Moving A2A to the Linux Foundation sounds like a gift to the open-source community. But Linux Foundation stewardship doesn't automatically mean community control. Who makes the breaking changes? What's the RFC process? The docs don't say yet.
Adoption isn't guaranteed. Microsoft has its own agent communication story. AWS has just launched AgentCore. The "one protocol to rule them all" moment hasn't happened. A2A being integrated into AutoGen and Semantic Kernel (Microsoft frameworks!) is encouraging — but enterprises won't abandon their existing stacks for a v1.2 spec without a lot more momentum.
The complexity cliff is real. Multi-agent coordination sounds elegant in keynote demos. In practice, debugging why Agent B misinterpreted a handoff from Agent A — across different models, different clouds, different retry policies — is going to be genuinely painful. The observability tooling (Google's Agent Observability, ADK tracing) is early.
The Bigger Picture
NEXT '26 was framed around one phrase: "The era of the pilot is over. The era of the agent is here."
Kurian was talking about enterprise adoption. But for developers, I'd reframe it: the era of building single-model, single-platform AI features is giving way to something messier and more powerful — networks of agents, built by different teams, running on different infrastructure, coordinating in real time.
The Gemini Enterprise Agent Platform is Google's bet on how you'll build those networks. A2A and MCP are the invisible foundation that makes it possible to build them without surrendering to a single vendor.
That's the story I'll be watching unfold. And it started quietly, in the keynote, in a slide most people skimmed past.
Top comments (0)