Google’s Gemini Enterprise Agent Platform makes Vertex AI the agent factory

Google Cloud has launched Gemini Enterprise Agent Platform, and the important line is not the usual “build agents faster” language. It is the platform move underneath it. In Google Cloud’s announcement of Gemini Enterprise Agent Platform, the company describes the product as the evolution of Vertex AI, combining model selection, model building, and agent building with newer pieces for integration, DevOps, orchestration, and security. Then it says the quiet part plainly: going forward, all Vertex AI services and roadmap evolutions will be delivered through Agent Platform rather than as a standalone service.

That is not a feature launch. That is a migration of the center of gravity. Vertex AI was the place enterprises went to manage models. Gemini Enterprise Agent Platform is the place Google wants them to manage work performed by agents. Models are still there, but they are no longer the main unit of product gravity. The unit is now the governed agent: something with a runtime, a memory, a set of tools, an identity, an approval path, an evaluation loop, and a trail of evidence when it breaks something at 2:13 p.m.

The real story is that enterprise agents are becoming less like chatbots and more like junior software services with ambitious LinkedIn profiles. A chatbot can be annoying. An agent connected to internal systems can be expensive, leaky, or operationally dangerous. That changes the buyer question. “Which model is smartest?” becomes “Which platform can let this thing act without turning security, compliance, and operations into a weekly séance?” Google’s answer is to put the whole agent lifecycle under one roof.

The platform is organized around four verbs: build, scale, govern, and optimize. The build layer includes Agent Studio for low-code work and an upgraded Agent Development Kit for code-first teams, with graph-based multi-agent patterns and workspaces for isolated command and file execution. The scale layer includes a reworked Agent Runtime, long-running agents that can maintain state for days, Agent Sessions, bidirectional streaming, and Memory Bank for persistent context. The govern layer is where the enterprise sale lives: Agent Identity, Agent Registry, Agent Gateway, Model Armor, anomaly detection, threat detection, and a security dashboard tied into Security Command Center. The optimize layer adds simulation, evaluation, observability, execution traces, and an Agent Optimizer that clusters real-world failures and suggests instruction refinements.

This is a lot of nouns. Some of them will surely age into product pages that require a second monitor and patience. But the bundle is pointed at a real problem. Most companies do not fail with agents because nobody can produce a demo. They fail because the demo does not survive contact with permissions, background jobs, tool access, stale context, audit requirements, and users who keep asking reasonable things in unreasonable order. The missing layer is not another animated sparkle button. It is operational control.

Google is also making the model-routing pitch explicit. Agent Platform provides access to more than 200 models through Model Garden, including Google’s Gemini 3.1 Pro, Gemini 3.1 Flash Image, Lyria 3, and Gemma 4, plus third-party options such as Anthropic’s Claude Opus, Sonnet, and Haiku. That is strategically important because enterprises do not want agent infrastructure that assumes one model wins forever. They want a stable control plane that can swap models as price, latency, licensing, privacy, and capability shift. Features are easier to demo than procurement flexibility. That does not make flexibility less important.

The customer examples in Google’s post are useful mostly because they show the range of jobs Google wants the platform to absorb: Color Health using ADK and Agent Runtime for a virtual cancer clinic workflow, Comcast rebuilding Xfinity Assistant around a multi-agent architecture, Payhawk using Memory Bank so finance agents can remember user-specific constraints, PayPal tying Agent Platform to agent payments, and L’Oréal talking about agents connected through Model Context Protocol to internal systems of record. These are not “write me a poem” cases. They are agent-as-interface-to-business-process cases. That is where governance stops being a checkbox and starts being the product.

The catch is that an integrated platform also becomes a gravity well. If your agents use Google’s runtime, memory, gateway, registry, identity, evaluation, observability, and security surfaces, moving later is not a one-line model swap. That may be fine. Most enterprises already choose gravity wells; they just prefer the ones with support contracts. But builders should be honest about the architecture trade. A managed agent platform can reduce the number of custom systems you have to invent. It can also make your agent stack inherit the roadmap, pricing, and abstraction choices of the vendor operating the platform.

The practical move for teams is not to “adopt agents” because a large cloud company has printed the phrase in fresh ink. It is to map where agent work is already slipping from experiment into operations. If agents are touching customer support, sales research, finance workflows, document review, infrastructure tasks, or internal knowledge systems, the questions are immediate: what identity does the agent use, which tools can it call, where is its memory stored, how are failures reviewed, how are prompts and policies versioned, and who can prove what happened after the fact? If those answers currently live in a spreadsheet and three heroic engineers’ heads, a platform like this is worth testing.

The thing to watch next is whether Google can make the governed path feel faster than the improvised path. Enterprises say they want guardrails, but builders route around platforms that turn every experiment into a ticket queue. Agent Studio, ADK export, prebuilt Agent Garden templates, batch and event-driven agents, and code-agent interfaces are all attempts to keep the path from becoming bureaucratic paste. The winning platform will not be the one with the thickest governance brochure. It will be the one where good governance is the fastest way to ship.

For Useful Machines readers, the takeaway is simple: Google is not merely launching another agent builder. It is declaring that the next phase of Vertex AI is agent infrastructure, not model hosting. If you are buying enterprise AI this year, pay less attention to which vendor has the most impressive stage demo and more attention to who can give every agent an identity, a memory boundary, a runtime, a trace, a test harness, and a kill switch. The future of agents is not autonomy theater. It is boring control surfaces, because boring control surfaces are what let software do real work without making everyone nervous.