Google Announces Gemini Enterprise Agent Platform at Cloud Next '26 in Las Vegas

Google has announced at the Cloud Next '26 conference in Las Vegas a new platform that entirely replaces Vertex AI as the central environment for building and deploying AI agents in enterprises. Gemini Enterprise Agent Platform has been available since April 22, 2026, and from that point takes over all further updates and development plans for Vertex AI.

Key Highlights

Google announced Gemini Enterprise Agent Platform as Vertex AI's successor — all future updates to this platform will go exclusively to the new environment
The platform provides access to over 200 models through Model Garden, including Google's own models (Gemini 3.1 Pro, Lyria 3, Gemma 4) and third-party models, among them Claude Opus 4.7 from Anthropic
The four pillars of the platform's architecture are: building agents, scaling to production, enterprise-grade security management, and optimization and observability tooling
Among the first production deployments are Comcast, PayPal, L'Oréal, and Geotab
The Agent Development Kit (ADK) processes over six trillion tokens per month

Vertex AI Is Retiring. The Platform Takes Over.

Vertex AI, which Google launched in 2021 as an environment for training, fine-tuning, and deploying AI models, is now being absorbed into the broader structure of the new platform. This is not a simple rebranding — all Vertex AI services and the subsequent stages of its development roadmap will from now on be delivered exclusively within Agent Platform, not as a separate service.

The announcement was made during CEO Sundar Pichai's keynote address. The new platform is intended to provide a secure, full-stack infrastructure connecting data, people, and organizational goals — "a command center for the agentic enterprise," as Pichai put it. Existing APIs remain backwards compatible, and current Vertex AI customers see the new brand directly in the console without any manual migration required — this follows from documentation published by Google.

Four Pillars: Build, Scale, Govern, Optimize

The platform's architecture rests on four clearly delineated layers, each addressing a different stage of the agent lifecycle.

The first layer, dedicated to building agents, encompasses two core tools: Agent Studio (a low-code visual environment designed for non-technical employees) and the Agent Development Kit — a comprehensive environment for engineering teams. The ADK has undergone significant modernization: a new graph-based orchestration mechanism allows designing networks of collaborating agents with explicit dependency logic, while a multimodal streaming mode handles both audio and video in real time.

The second layer handles the long-term operation of agents in production environments. Agent Runtime ensures agents remain continuously operational for days or weeks. Memory Bank enables state to be maintained across an entire long sequence of steps, while Memory Profiles allow persistent contextual profiles to be built for individual users and sessions.

The third layer addresses the problem most frequently cited as the brake on enterprise AI deployments — the lack of control and auditability. Agent Identity assigns each agent a cryptographic identifier. Agent Registry creates a catalog of approved agents available to the entire organization. Agent Gateway enforces security policies, protecting against threats including tool poisoning and data leakage. This is complemented by the Model Armor layer — security at the model level.

The fourth layer brings together tools for observability and improvement: Agent Simulation for pre-deployment testing, Agent Evaluation for ongoing behavioral assessment, and Agent Optimizer, which — according to Google's description — uses a combined LLM-as-judge model to evaluate response quality.

Model Garden: Over 200 Models, Open Ecosystems

The platform provides access to over 200 models through Model Garden. The list includes Google's latest proprietary models — Gemini 3.1 Pro, Gemini 3.1 Flash Image, and Lyria 3 — alongside open Gemma 4 models and models from third-party providers: Claude Opus 4.7, Sonnet, and Haiku from Anthropic. The platform supports the A2A (Agent-to-Agent) and MCP (Model Context Protocol) protocols, enabling integration with third-party systems and tools without custom programming.

In parallel, Google updated the Gemini Enterprise application — the interface aimed at knowledge workers. New features include Agent Designer for creating automations in natural language, an agent inbox for monitoring long-running processes, and support for Google Workspace, Microsoft 365, and third-party systems via BYO-MCP connectors.

Early Deployments: Comcast, PayPal, L'Oréal

Several large organizations have already launched production deployments on the new platform. Comcast rebuilt its Xfinity assistant using ADK, transitioning from simple script-based automation to generative customer service with the ability to resolve issues autonomously. PayPal uses the platform to shorten financial application processing times and for autonomous commercial interactions via the Agent Payment Protocol. L'Oréal built an internal agent platform connected to the company's data systems.

Competitive Context: AWS vs. Google

The new platform enters a market where a clear philosophical divide is emerging. As analyses published by VentureBeat indicate, Google and Amazon Web Services are taking fundamentally different approaches to agent management. Google bets on a central system-level control layer — Gemini Enterprise Agent Platform acts as the governance and audit mechanism for the entire agent environment. AWS, on the other hand, focuses on acceleration at the execution level — Bedrock AgentCore simplifies running and deploying agents while delegating more decisions to the project level. Both approaches address real needs, but target different pain points: Google targets control over an agent fleet, AWS targets speed to production.

Microsoft with Azure AI Foundry and Amazon with Bedrock AgentCore are the direct competitors to Google's offering. Positioning against them is clearly visible in the distribution model: Google points to full-stack ownership — chips, models, cloud runtime, and workspace — as an advantage that no competitor can fully replicate. More than half of Google's investment in machine learning compute in 2026 is expected to go to cloud products.

Why Does This Matter?

The closure of Vertex AI as a standalone product is a strategic signal that goes beyond a mere name change. Google is no longer building a platform for AI developers — it is building operational infrastructure for enterprises that need to manage hundreds or thousands of active agents simultaneously. That is a fundamentally different problem from launching a single model or a single chatbot.

The governance layer — Agent Identity, Registry, and Gateway — addresses the question that has stalled enterprise deployments: who oversees what an agent does on behalf of the organization, and how? Without cryptographic agent identities, a central catalog of approved tools, and automatically enforced security policies, every agent deployment in a regulated environment demands costly manual auditing.

Equally significant is the native support for Claude Opus 4.7 from Anthropic directly in Model Garden. This signals that Google prefers to build an open model ecosystem rather than a closed vertical that would require customers to abandon their preferred providers. If this approach holds, the platform may attract organizations that do not want to become dependent on a single model vendor.

What's Next?

New Gemini Enterprise capabilities are set to reach users gradually over the coming months — Google has not provided a detailed timeline for individual features
The key test will be the effectiveness of long-running agents in regulated environments (finance, healthcare, law), where auditability requirements are highest
It is worth watching how the platform handles the problem of state drift in long-running agents — a technical issue identified by VentureBeat as a new class of failure in agentic systems