What is Gemini Enterprise Agent Platform? (May 2026)
What is Gemini Enterprise Agent Platform? (May 2026)
Google’s Gemini Enterprise Agent Platform is the agent runtime announced at Cloud Next 2026 on April 22. It supersedes Vertex AI Agents as Google’s primary enterprise agent story, adds four governance components (Registry, Gateway, Identity, Sandbox), and is one half of the May 2026 enterprise agent war alongside Microsoft Agent 365 and OpenAI Workspace Agents.
Last verified: May 12, 2026
At a glance
| Property | Value |
|---|---|
| Vendor | Google Cloud |
| Announced | April 22, 2026 (Cloud Next 2026) |
| Status | GA on Vertex AI |
| Replaces / extends | Vertex AI Agents |
| Default model | Gemini 3.1 Pro |
| Also supports | Gemini 3.1 Flash Image (Nano Banana 2), Lyria 3, Claude Opus / Sonnet / Haiku |
| Orchestration | Central Gemini coordinator → specialized sub-agents |
| Long-running agents | Yes — days-long execution supported |
| Open standards | Agent2Agent (A2A) protocol |
| Native integrations | Workspace (Gmail, Calendar, Drive, Docs, Sheets), NotebookLM Enterprise, Gemini Code Assist |
| Marketplace integrations | Jira, Salesforce, ServiceNow, Atlassian, Box, Oracle, Workday |
The four governance components
Most of what is genuinely new in the platform is in the control plane. Google is treating “agent sprawl” as the next shadow-IT problem and positioning Registry/Gateway/Identity/Sandbox as the answer.
1. Agent Registry
A central catalog of approved tools, skills, and agents inside an organization. Only registered items can be discovered or invoked by other agents. Think of it as an internal app store for agentic capabilities — with approval flow, ownership, version history, and access control. It is what gives a security or platform team something to govern.
2. Agent Gateway
The runtime policy enforcement layer between agents and the tools they call. Every tool call passes through the Gateway. Functions as “air-traffic control”:
- Authentication and authorization for tool access.
- Prompt-injection and data-leakage defense (Model Armor).
- Audit logging of every tool invocation.
- Quota, rate limit, and cost guardrails.
The Gateway is the choke point that makes Registry policies actually enforceable.
3. Agent Identity
Each agent gets a unique cryptographic ID. Tool calls, audit logs, and inter-agent messages are attributable to a specific agent rather than to a service account shared by many agents. This is the foundation for incident forensics, fine-grained access control, and ultimately for letting third-party agents act on behalf of an organization without impersonation risk.
4. Agent Sandbox
A hardened, isolated environment for executing model-generated code and computer-use tasks. Browser automation, code execution, and arbitrary tool invocations run inside the sandbox rather than directly on host systems. This is structurally similar to the sandboxing layer that emerged after the Trustfall AI-coding-agent attack class became public in late April 2026 — the industry-wide answer to “don’t let an agent’s generated code own your VM.”
Multi-agent orchestration
A central Gemini model acts as the coordinator. The user (or another agent) gives a natural-language prompt, and the coordinator decomposes the work, dispatches sub-tasks to specialized sub-agents, and integrates the results. The Agent Development Kit (ADK) exposes a graph-based framework for organizing sub-agents — nodes are agents, edges are message and data flows, and the runtime handles retries, fan-out, and state.
This is structurally similar to LangGraph but managed, governed, and tied to Workspace identity by default.
Long-running agents
The platform explicitly supports agents that run autonomously for days. Background dispatching, scheduled execution, and persistent state across runs are first-class. This puts it in the same conceptual neighborhood as Claude Managed Agents (Anthropic) and OpenAI Workspace Agents — the three platforms converging on the “agent that runs while you sleep” use case.
Models on offer
The platform supports a deliberately mixed model lineup:
- Gemini 3.1 Pro — the default reasoning model.
- Gemini 3.1 Flash Image (Nano Banana 2) — fast vision/image workloads.
- Lyria 3 — Google’s audio/music model.
- Anthropic Claude Opus 4.7, Sonnet 4.6, Haiku 4 — full Claude family available, contracted through Google Cloud.
This is significant: Google is shipping a first-party Claude integration in the same product where it ships Gemini. The buyer pitch is “pick the best model per agent, not per vendor.”
Workspace and marketplace integration
Native, deep integration with Google Workspace — Gmail, Calendar, Drive, Docs, Sheets, Meet. Agents can read and act on Workspace artifacts under the agent’s Identity, governed by the Gateway. The Workspace Marketplace and Agentspace add third-party tools — Jira, Salesforce, ServiceNow, Atlassian, Box, Oracle, Workday — through governed connectors.
The Gemini Enterprise App is the user-facing hub: employees see all approved organizational agents in one place, including specialized ones like Data Insights, NotebookLM Enterprise, and Gemini Code Assist.
How it compares to the alternatives
| Capability | Gemini Enterprise Agent Platform | OpenAI Workspace Agents | Microsoft Agent 365 |
|---|---|---|---|
| Agent runtime | ✅ Managed, multi-agent graph | ✅ Codex-powered, scheduled | Control plane (uses Copilot Studio + others) |
| Default model | Gemini 3.1 Pro | GPT-5.5 family | GPT-5.5 Thinking + others |
| Non-Google models supported | Claude Opus/Sonnet/Haiku | OpenAI only | Multiple via Foundry / partners |
| Governance components | Registry, Gateway, Identity, Sandbox | Compliance API, permission controls | Observability + approval + Shadow AI detection |
| Native productivity stack | Google Workspace | Slack + Drive + MS apps + Salesforce | Microsoft 365 (deepest of the three) |
| Long-running agents | ✅ Days | ✅ Scheduled | Via underlying platforms |
| A2A interop | ✅ | Roadmap | Via partners |
| GA | ✅ April 22, 2026 | Research preview | ✅ May 1, 2026 |
What it doesn’t do
- Not a model. It hosts models; it isn’t one.
- Not free. Pricing is per-token plus runtime/storage on Vertex AI.
- Not Workspace itself. It integrates with Workspace; it doesn’t replace it.
- Not fully open. Open standards (A2A) are supported, but the Registry/Gateway/Identity stack is Google-managed.
What to watch next
- Gemini 4 timing. Google has signaled enhanced reasoning and a massive context window for the next generation.
- A2A protocol adoption. Whether Anthropic and OpenAI ship A2A compatibility in their own runtimes.
- Sandbox feature parity. Whether the Agent Sandbox catches up with Anthropic’s Claude Managed Agents sandbox on Trustfall-class defenses.
- Pricing visibility. Long-running agents at “days” durations need predictable cost models — watch for committed-use discounts and per-agent caps.
Sources
- Google Cloud Blog, “Introducing Gemini Enterprise Agent Platform”
- blog.google, “Gemini Enterprise Agent Platform” announcement
- Virtualization Review, “Google Cloud Next ‘26: Gemini Enterprise Agent Platform leads AI-centric news”
- MindStudio, “Gemini Enterprise Agent Platform business automation”
- Google Cloud, “Gemini Enterprise / Agents” product page
Related reading
- Gemini Enterprise vs Agent 365 vs Bedrock Managed Agents
- OpenAI Workspace Agents vs Gemini Enterprise Agent vs Agent 365
- What is Gemini 3.1 Pro
- Microsoft Agent 365 GA vs Amazon Quick vs Google Workspace Studio
- What is Claude Managed Agents
Last verified: May 12, 2026.