Google used the opening day of Cloud Next 2026 in Las Vegas to reframe how it wants enterprises to build with AI — folding Vertex AI, its flagship model-building service, into a broader package called the Gemini Enterprise Agent Platform. Announced on April 22, the platform is designed to handle the full lifecycle of autonomous agents, from development through deployment, governance, and continuous optimization.
The move signals that Google sees the next wave of enterprise AI revenue not in raw model APIs, but in the control plane around multi-step, long-running agents — the same category being chased by OpenAI's workspace agents and Anthropic's Claude-based enterprise tooling.
Four layers: Build, Scale, Govern, Optimize
The platform is organized around four pillars. For building agents, Google is shipping Agent Studio, a low-code visual interface, alongside the code-first Agent Development Kit (ADK) and an Agent Garden of pre-built templates for tasks like code modernization, financial analysis, and invoice processing. A hardened Agent Sandbox and secure Workspaces let agents run bash commands and manipulate files without compromising customer environments.
To scale agents in production, Google is introducing an Agent Runtime with what it describes as sub-second cold starts, support for multi-day autonomous workflows, and an Agent Memory Bank that generates long-term memories tied to user profiles. Bidirectional WebSocket streaming and persistent session IDs round out the runtime.
The governance layer is where Google is trying to differentiate. Every agent gets a unique cryptographic Agent Identity for auditable action trails. An Agent Registry acts as a central library of approved tools and skills, while the Agent Gateway functions as what Google calls "air traffic control" for the agent ecosystem. Model Armor guards against prompt injection and data leakage, and an Agent Security dashboard plugs into Security Command Center.
For optimization, Agent Simulation runs tests against synthetic users and virtualized tools, while Agent Observability provides visual traces of reasoning steps and Agent Optimizer clusters failures into suggested system-instruction fixes.
Model choice and early customers
Google is emphasizing model openness. The platform exposes more than 200 models, including its own Gemini 3.1 Pro, Gemini 3.1 Flash Image, Lyria 3, and the open Gemma 4 family — plus Anthropic's Claude Opus, Sonnet, and Haiku. Support for the Model Context Protocol and Google's Agent Payment Protocol (AP2) is also baked in.
Launch customers include Comcast, which is rebuilding its Xfinity Assistant on ADK; L'Oréal, whose Group CIO Etienne Bertin cited the platform as the foundation for its Beauty Tech Agentic Platform; PayPal, which is using AP2 for agent-based commerce; and Payhawk, whose Financial Controller Agent reportedly cuts expense-submission time in half.
Implications
By explicitly stating that all Vertex AI roadmap work will flow through Agent Platform, Google is betting that agents — not models — are the enterprise product. That aligns the company's cloud pitch with a broader industry shift toward agentic AI replacing enterprise automation in 2026, and intensifies competitive pressure on rivals selling agent tooling piecemeal.
For buyers, the harder question is not which platform to pick but how quickly any of them pay back. This breakdown of AI agent time-to-ROI use case benchmarks for 2026 maps typical payback windows by use case, and the consolidation trend Google is pushing is itself a response to a growing pain point: AI sprawl, the 94% problem strangling enterprise AI in 2026, which makes a single governed control plane more attractive than another standalone tool.
