I recently attended Google Cloud Next 2026, where Google laid out a significantly retooled agent strategy anchored on the new Gemini Enterprise Agent Platform — the rebrand of Vertex AI that absorbs Agentspace and the Agent Development Kit into a single offering. Initially, I was underwhelmed by the announcement, but after digging deeper, I changed my view. Google, along with a small set of competitors, is building what I think the market actually needs: an agentic control plane. And Google has enough distinct capability inside this platform to put itself in serious contention for category leadership.

Why We Need an Agentic Control Plane

I have been researching agentic control planes since early 2025, working from the premise that enterprise AI cannot scale without a central place to govern, optimize, and observe the agents it produces. The first generation of these platforms — call them gen-one control planes — was foundational. They standardized development practices, allowed for model choice, and added baseline security and observability. That helped organizations contain costs and apply some governance. But they were aimed almost entirely at professional developers and infrastructure architects, which made sense because at that time those groups were the only people who could realistically build agents.

Since then, the rapid growth of no-code agent tooling has forced vendors to evolve. Gen-two control planes have to support a much broader population of builders and the agents they ship. That is where Google sits today, and it is why I think this new announcement places Google in a small group of vendors with a credible shot at mainstream agent share. Whether this generation finally solves the historical pathologies of citizen-developer platforms — sprawl, security drift, unmaintained sprawl-as-tech-debt — is the more interesting open question.

Is the Gemini Enterprise Agent Platform Actually New?

Sitting through the keynotes, I had a debate going in my own head. On the surface, the platform looks like a repackaging of existing Google assets — Vertex AI, Agentspace, and the Agent Development Kit rolled into one storefront. There is enough new foundational substrate, however, to justify treating this as a new product rather than a mere relabeling. Some of the additions close gaps Google had against competitors; memory and observability are two examples. Others are areas where Google has carved out distinct positioning. Here are the three additions I find most consequential.

Agent Skills Repository. Skills have become an important construct for agents over the past year. They let a builder inject capability on demand instead of stuffing every possible instruction into a long, token-hungry prompt. Until now, however, skills have largely been a personal artifact — most familiar to anyone who has worked inside Claude Desktop, where skills live with the individual user. Google is publishing an official Agent Skills Repository covering products like BigQuery, GKE, Cloud Run, and the Gemini API, which moves skills from a personal tool into a governed, enterprise-grade primitive on par with models and MCP tools. That elevation is the right move, and it is the part of the announcement most likely to be copied by competitors.

Agent Simulation. Agent testing today is mostly two steps: confirm the agent runs, then turn it loose on real users and see what breaks. That is operationally costly. Builders need testers — who are reluctant to be guinea pigs — and the diagnostic loop after a failure is slow and expensive. Agent Simulation generates synthetic, multi-step user interactions in a sandboxed environment so a builder can stress-test an agent before deployment, with automatic scoring across task success and safety. If it works as advertised, this should compress quality cycles meaningfully and is one of the more practical pieces of the announcement.

Agent Anomaly Detection. The industry has made real progress on agent identity, access control, and on-behalf-of authorization. Runtime threats such as prompt injection are the harder problem. Agent Anomaly Detection uses an LLM-judge approach to flag suspicious behavior while an agent is running — including tool misuse, unauthorized data access, reasoning drift, and the more familiar offenses like reverse shells. When the system detects an anomaly, it can intervene before the action completes. Whether this holds up against adversarial prompts in production is the test that matters; published research still shows attack success rates above 85% against state-of-the-art defenses, so claims of “stopped before the attack happens” need to be earned, not asserted.

Are These Platforms Ready for Production Use?

If your company is moving past pilots and into early production agents, the answer is largely yes. The mental model that helps here: An agent is not the application, but rather a workflow that steers an AI model toward a useful outcome. In an enterprise context, “useful” means consistent, secure, and cost-optimized. Google and its main competitors have closed real ground on those three attributes over the past year. That said, there are still a handful of issues buyers should weigh, regardless of whether they are evaluating Google, Microsoft, or Salesforce.

The Claude-ification of the Agent UX. Every one of these platforms supports a wide range of front-ends: desktop productivity suites, custom applications, embedded experiences. Each is also being marketed alongside a built-for-purpose agent desktop app, and after reviewing them through the spring, those apps have converged on something that looks an awful lot like Claude Desktop. That is not a bad thing for users, but it does mean that UX is no longer a meaningful differentiator. Buyers should not let the demo carry the decision.

Enterprise Selection Will be Decided by the Details. If UX is not the differentiator, what is? Two things, in my view. The first is your existing application footprint. If you are already invested in Microsoft 365 and Entra ID, the rational starting point is the incumbent; the same logic applies for a Salesforce or ServiceNow shop. The second differentiator, more important over a longer horizon, is full-stack economics. Google is making the argument that its tightly coupled stack — TPUs, Gemini multimodal models, with Workspace on top — produces a better cost-and-performance curve than a multi-vendor assembly. The argument is plausible because Google claims to be the only hyperscaler that owns every layer, but that is a point-in-time statement given that Microsoft will also have its custom accelerators in production soon; whether the math actually plays out in customer P&Ls is something I want to see proven across more workloads before I take it as gospel. A Microsoft or Salesforce buyer will reasonably push back on the framing.

Will AI Finally Make Citizen Development Governable? I have been building on no-code platforms for more than thirty years (Lotus Notes 2.1 on OS/2, thank you very much), and no platform I have used has fully solved citizen-developer governance. Sprawl, enterprise scaling, and accumulated tech debt eventually consume every successful citizen-developer platform from the inside. What is different this time is that AI itself can be used to police what AI builders are doing. The observability surfaces inside the Gemini Enterprise Agent Platform, combined with maturing third-party tooling, give me cautious optimism that this generation can avoid the trap. I would not bet the company on it yet, but the architectural ingredients are in place for the first time.

What Google Still Needs to Prove

A few things would move my conviction about Google from “serious contender” to “category leader.” First, I’d like to see real production proof of A2A interoperability with non-Google agents; standards on paper are easy, and the Linux Foundation governance helps, but cross-vendor agent handoffs in customer production are the only test that matters. Second, let’s see governed defaults for citizen-built agents. The Agent Registry framing is right, but most enterprises will need opinionated, ship-ready guardrails rather than a configuration menu, especially in early days of use. Third, Google needs to show sustained model leadership. The full-stack economics argument only holds if Gemini stays at or near the frontier; if it slips, the buyers Google is courting will route workloads to whoever is leading that quarter.

Net-net, I agree with my colleague Patrick Moorhead’s sentiments on Google and its recent agent announcements — and that includes the Gemini Enterprise Agent Platform. In other words, I believe that this is the most coherent statement Google has made yet on how enterprises should build, govern, and run agents at scale. It does not yet settle the category, and Microsoft and Salesforce will not concede the buyer-of-record relationship without a fight. But Google is now playing in the right game with the right pieces, and that is a meaningful change from where the company was this time last year.