Google Cloud Next '26: The AI Announcements That Matter

Google Cloud Next '26 dropped a year's worth of agent announcements into one keynote, and most of the coverage read like a press release. If you actually run workloads on Google Cloud, the question isn't "what got announced" — it's "what changes my architecture this quarter, and what's a slide that won't matter until next year."

Here's the practitioner cut. I'm separating the things that move your roadmap from the keynote noise.

The headline: Vertex AI became the Gemini Enterprise Agent Platform

The biggest structural change is a rename with teeth. Vertex AI is now positioned as the Gemini Enterprise Agent Platform, described by Google as the evolution of Vertex AI — future services and roadmap land there rather than as standalone Vertex offerings.

Don't shrug this off as branding. When a platform absorbs the thing you already build on, you inherit a migration story whether you wanted one or not. Your existing endpoints, IAM bindings, and pipelines keep running, but new capabilities increasingly show up under the new umbrella. The practical move: inventory what you depend on in Vertex AI today and watch where the parity gaps land over the next two quarters.

The agent layer on top is where the actual product is. The Gemini Enterprise app picked up an Agent Designer (no-code, trigger-based workflows), an Agent Inbox for monitoring what your agents are doing, and long-running agents that operate in the background instead of dying at the end of a request. There's also an Agent Studio, a lower-code interface for building agents with natural language.

The long-running, background-agent piece is the one that matters most. The jump from "chatbot that answers and forgets" to "agent that holds a task for hours and reports back" is the whole reason this category exists. If you've been hand-coding orchestration and state for that, this is the part to evaluate first.

Model Garden: 200+ models, including the ones you'd actually pick

Model Garden now hosts more than 200 models. The roster includes Google's own — Gemini 3.1 Pro and Gemini 3.1 Flash Image — alongside third-party models like Anthropic Claude and open models including Llama and Mistral.

The signal here is choice without leaving the platform. You can route to Claude for the reasoning-heavy step and a cheaper open model for bulk work, under one billing and governance layer. For teams that were stitching together multiple vendor SDKs, consolidating that into one model-access layer is a real reduction in glue code.

The honest caveat: "available in Model Garden" and "GA, supported, and priced the way you'd want" are not the same thing. Check the launch stage and the per-model terms before you design around any single one.

BigQuery AI: inference moves into SQL

This is the quiet announcement that I think has the longest tail. BigQuery added managed and SQL-native inference for open and third-party models — you can run open models with a couple of SQL statements, and BigQuery provisions and then releases the compute for you.

Deploy an open model with a single CREATE MODEL statement, run inference in SQL, then drop the model and BigQuery cleans up the associated Vertex resources. Idle compute gets released automatically to avoid surprise bills.

Why this matters: it collapses the gap between your data and your models. Analysts who already live in SQL can run inference where the data sits, without exporting to a separate serving stack. That's a genuinely different workflow, not a faster version of the old one. Several of these inference features shipped in Preview, so confirm the stage before you put it on a critical path.

Agent tooling: ADK and A2A grew up

Two pieces of plumbing matured enough to build on:

• Agent Development Kit (ADK) reached stable 1.0 releases across multiple languages. Stable is the word that matters — it's the difference between "interesting repo" and "thing you put in production."
• Agent2Agent (A2A) is now stewarded by the Linux Foundation and reported in production at a number of organizations. It's the protocol for agents handing tasks to each other across platform and org boundaries.

A2A and MCP are not competitors, despite how the takes read. MCP connects an agent to its tools and context; A2A connects agents to other agents. If your roadmap has multiple agents from multiple vendors that need to coordinate, A2A is the layer. If you're wiring one agent to tools and data, MCP is still the story. Most teams need both, eventually.

Workspace AI: useful, not the headline

Workspace got a no-code environment for building and deploying agents across Gmail, Docs, Sheets, and Meet, connecting out to services like Salesforce and Jira, with "Ask Gemini in Chat" spread across the core apps.

This is real and your non-engineering colleagues will use it. But for a platform team, it's downstream of the agent platform decisions above. Treat it as adoption surface, not architecture.

What actually changes your roadmap

Strip the keynote down and three things deserve a calendar entry:

1. The Vertex AI → Gemini Enterprise Agent Platform shift. Plan for it now; it's the path your future workloads take.
2. SQL-native inference in BigQuery. If your team lives in BigQuery, this changes who can ship model-backed features.
3. Stable ADK + A2A. If you're building multi-agent systems, the plumbing is finally production-grade.

Everything else — the TPU generations, the security agents, the Workspace polish — is real, but it's either infrastructure you'll consume without thinking about, or features you'll adopt without re-architecting.

One discipline to carry out of this: a lot of these capabilities are Preview, and Google reshuffled enough product names this cycle that yesterday's blog post may already use the old branding. Confirm the launch stage and the current name of anything before you build on it. Being early is an advantage; being early on something that's still Preview and might get renamed is just a support ticket waiting to happen.