Revolgy blog

The definitive guide to Gemini Enterprise Agent Platform (GEAP)

Written by Muskan Goel | June 22, 2026

For the past three years, the enterprise AI conversation has been dominated by one thing: chatbots. Virtual assistants. Copilots. AI that answers questions, drafts emails, and summarizes. Useful, but incremental. The AI was still waiting for you to tell it what to do next.

That model is changing, and the change is structural.

In April 2026, Google rebranded Vertex AI and launched the Gemini Enterprise Agent Platform at Google Cloud Next '26. The timing was not accidental.

Microsoft had already embedded Copilot across its enterprise stack. Amazon had launched Bedrock AgentCore. Anthropic had made Managed Agents generally available. Every major cloud vendor was moving in the same direction.

The distinction matters more than it sounds. A copilot helps you do a task. You stay in the loop, you make the decisions, the AI handles parts of the execution. An agent is different.

You give it a goal, and it figures out the steps, selects the tools, sequences the actions, and delivers an output without you managing each step along the way.

The difference is autonomy, and autonomy at the scale enterprises actually need it is a much harder engineering problem than building a better chatbot.

That is the problem Gemini Enterprise Agent Platform is built to solve. This guide covers everything you need to know about the platform: what it is, how it works, where it genuinely delivers, and what it takes to deploy it in a real engineering environment.

 

Jump to section:

 

What is Gemini Enterprise Agent Platform?

 


 

Gemini Enterprise Agent Platform is Google Cloud’s unified infrastructure for building, deploying, governing, and optimizing AI agents at enterprise scale.

Announced at Google Cloud Next '26, it is the direct evolution of Vertex AI, consolidating what was a fragmented set of ML tools into a single platform purpose-built for the agent era.

It is not a chatbot tool or a workflow automation product. It is closer to what Kubernetes is for containers: a managed foundation that handles the hard infrastructure problems so engineering teams can focus on what their agents actually do.

The platform is built for technical teams: ML engineers, platform engineers, cloud architects. It requires Google Cloud access, an understanding of agent architecture concepts, and engineering resources to build and govern agents properly.

The platform became generally available on the same day it was announced and is accessible through the Google Cloud console.

Vertex AI vs Gemini Enterprise Agent Platform

Vertex AI was Google Cloud’s machine learning platform, launched in 2021. It was where teams built, trained, deployed, and monitored ML models. Over time, it expanded to include conversational AI agents, model fine-tuning, a model garden, and data pipelines.

In April 2024, Google launched Vertex AI Agent Builder within it, the first dedicated tool for building production AI agents on the platform. The problem was that each of these capabilities lived in a slightly different part of the console, required a different mental model, and had its own documentation, pricing, and support surface. Powerful, but fragmented.

By early 2026, the enterprise AI conversation had moved decisively toward agents. At Google Cloud Next ’26, Google announced that Vertex AI would be consolidated into the Gemini Enterprise Agent Platform. Google was explicit: all Vertex AI services and roadmap evolutions will be delivered exclusively through Agent Platform going forward.

Vertex AI as a standalone product ceases to exist.

What actually changed with Vertex AI

The consolidation brought an entirely new layer of capabilities that Vertex AI never had: managed agent runtimes, persistent memory, cryptographic agent identity, centralized governance, and production-grade optimization tooling. These are covered in full in the Four Pillars section below.

Existing deployments do not require urgent migration. Google has maintained backward compatibility on existing APIs, the console location is unchanged, and the Model Garden remains intact. The new capabilities are simply available now under the Agent Platform brand when teams are ready to use them.

Full documentation is at docs.cloud.google.com/gemini-enterprise-agent-platform.

4 pillars of Gemini Enterprise Agent Platform

Google organized the platform around four functions: Build, Scale, Govern, and Optimize. Each pillar has specific tools under it.

 

Source: Google Cloud

 

Build

This is where agents are created. The platform offers two development paths depending on who is building and how much control they need.

  • Agent Studio is the low-code option. A visual interface where you design agent logic, connect data sources, and test prompts without writing code. It is where most teams prototype before moving to production, and it is where business-oriented builders who are not full engineers can contribute.
  • Agent Development Kit (ADK) is the code-first option. An open-source framework that gives developers precise control over agent behavior, tool use, and orchestration logic. It supports Python, Go, Java, and TypeScript. It is model-agnostic, meaning it works with Gemini, Claude, Llama, or any other model, and it deploys to any container or Kubernetes environment.
  • Data Agent Kit extends agent building directly into developer environments: VS Code, Claude Code, and Gemini CLI, so engineers do not need to leave their existing workflow to build agents.
  • Agent Garden is a library of pre-built agent templates for common enterprise use cases, including code modernization, financial analysis, invoice processing, and economic research. These serve as starting points, not finished products.
  • Model Garden gives access to over 200 foundation models from a single place. First-party models include Gemini 3.1 Pro, Gemini 3.1 Flash Image, Lyria 3, and Gemma 4. Third-party models available include Anthropic’s Claude Opus, Sonnet, and Haiku, as well as Llama.

Scale

Building an agent is the easy part. Running it reliably in production, with persistent memory, across thousands of simultaneous users and multi-day workflows, is where most agent projects break down. This pillar addresses that.

  • Agent Runtime is the managed execution environment. It handles autoscaling, sub-second cold starts, and lets teams provision new agents in seconds. It supports long-running agents that operate autonomously for days at a time, which is necessary for workflows like sales prospecting sequences, compliance monitoring, or deep research tasks that cannot be completed in a single session.
  • Agent Memory Bank gives agents persistent memory across conversations. Without it, every session starts from zero. With it, an agent remembers a user’s constraints, preferences, and historical decisions across interactions. Memory Profiles allow agents to retrieve high-accuracy context with low latency at scale.
  • Agent Sessions manage state within a single conversation. Custom Session IDs let teams map sessions directly to their internal database and CRM records.
  • Batch and event-driven agents connect to BigQuery and Pub/Sub, enabling agents to run large asynchronous tasks like data analysis or content evaluation in the background without blocking other operations.
  • Bidirectional Streaming via WebSocket enables real-time audio and video processing for human-like responsiveness in live interactions.

Govern

Governance is where enterprise AI projects either earn trust or lose it. An agent that can take real actions in production systems, query internal data, invoke tools, and coordinate with other agents, needs the same level of control and auditability that any other enterprise system would require.

  • Agent Identity assigns every agent a unique cryptographic ID. Every action an agent takes is mapped back to that identity and to the authorization policies governing it, creating a complete audit trail.
  • Agent Registry is the central library of every agent, tool, and skill across the organization. It controls what is available to whom and prevents the proliferation of ungoverned agents across departments.
  • Agent Gateway acts as the enforcement point for all agent connections. It controls which agent can connect to which system, what tools it can invoke, which user authorized the action, and what policies apply before any operation is allowed. It also enforces consistent security policies across agents built on Agent Platform and agents sourced from third-party partners.
  • Model Armor sits at the gateway level and provides runtime protection against prompt injection attacks and data leakage before requests reach the model.
  • Agent Anomaly Detection uses statistical models and an LLM-as-a-judge framework to flag unusual agent reasoning in real time. Agent Threat Detection identifies malicious activity, including reverse shells and connections to known bad IP addresses. Both feed into the Agent Security Dashboard, powered by Security Command Center, which unifies threat detection, risk analysis, and vulnerability scanning across the agent fleet.

Optimize

An agent that performs well in testing can still drift in production. This pillar provides the tooling to catch that drift and fix it systematically.

  • Agent Simulation tests agents before they ship by running them against synthetic human-like conversations and virtualized tools in a controlled environment. Each run is automatically scored on task success and safety across multi-step conversations.
  • Agent Evaluation scores agents continuously against live production traffic using multi-turn autoraters that evaluate the logic of an entire conversation, not just a single response.
  • Agent Observability provides full execution traces and a real-time view into agent reasoning, so teams can debug issues as they happen rather than after the fact.
  • Agent Optimizer removes the need to manually dig through failure logs. It automatically clusters real-world failures, identifies patterns, and suggests refined system instructions to improve agent accuracy over time.

How Gemini Enterprise Agent Platform works?

GEAP runs a coordinated network of specialized agents, each handling a discrete part of a task, all governed by platform-level controls at every step. Here’s how it works:

  • The orchestrator maps the task: A central agent built on ADK’s graph-based framework breaks the request into a workflow, assigning parts to specialized sub-agents, some running in parallel, some waiting on outputs from others before they proceed.
  • Agent Gateway controls every connection: Before any agent connects to a system, calls a tool, or passes data to another agent, Gateway checks authorization and enforces policy. Nothing moves without passing through this layer.
  • Agent Identity makes every action traceable: Every agent has a cryptographic ID. Every action it takes is logged against that identity and the authorization that permitted it, making auditing possible across an entire fleet.
  • Memory Bank keeps context alive: Within a session, agents share context through the session layer. Across sessions, Memory Bank retains what matters so agents do not start from zero on every interaction.
  • Agent Runtime handles execution: The managed runtime handles compute, scaling, and long-running workflows in the background. Engineering teams define what agents do; Runtime handles whether they stay running.
  • Observability records everything: Every step of every execution is traced. When something breaks, teams trace the reasoning path rather than guess.

Where is GEAP already being used?

The platform launched in April 2026, and companies across industries are already running production workloads on it. These are three documented examples from Google’s official announcement:

  • Payhawk built a Financial Controller Agent using Memory Bank that learns each user’s expense habits and constraints across sessions, cutting expense submission time by 50%.
  • Color Health used ADK and Agent Runtime to power a Virtual Cancer Clinic that autonomously checks patient screening eligibility, connects them to clinicians, and schedules appointments at scale.
  • L’Oreal built a proprietary Beauty Tech Agentic Platform using ADK connected via MCP to their internal data platform, moving from scripted workflow automation to agents that pursue outcomes autonomously.

Gemini Enterprise Agent Platform pricing

The platform operates on a pay-as-you-go model with no flat subscription fee. You are billed across multiple dimensions depending on what your agents actually consume.

What is billable, per Google’s official pricing page:

  • Model token consumption based on character input, image input, or custom training
  • Agent Runtime compute charged by vCPU hours and memory GiB hours
  • Storage at the same rates as Compute Engine and Cloud Storage
  • Management fees based on region and instances used
  • Agent Evaluation and Simulation runs charged per 1k characters of input and output

Google does not publish public pricing for the platform. Commercial terms are handled through enterprise contracts that vary based on usage volume, contract length, and support level. Organizations evaluating the platform should engage Google Cloud sales or a Premier Partner for a tailored quote.

For teams not already running workloads on Google Cloud, the unit economics are worth examining carefully before committing. The platform is most cost-efficient when your data already lives in Google Cloud infrastructure. Pulling data from external systems adds both cost and integration complexity.

New customers can access a free trial with up to $300 in Google Cloud credits at console.cloud.google.com/agent-platform/overview.

A practical roadmap to getting started with GEAP

Step 1: Set up prerequisites

You need a Google Cloud account with billing enabled. The platform is accessible from Google Console and new customers get $300 in free credits. Google Workspace is not mandatory but significantly expands what agents can access and act on.

Step 2: Define your first use case

Pick one workflow that is narrow, well-understood, and measurably valuable. Not “improve our operations with AI” but something specific enough that you will know within weeks whether it is working like “automatically triage incoming support tickets and route them to the right team with relevant context.”

Broad first use cases almost always fail. Narrow ones build confidence and surface real constraints before you scale.

Step 3: Choose your starting tool

Start with Agent Studio if you are prototyping or involving non-engineering stakeholders in design. Move to ADK when building for production or working with complex multi-agent workflows. Most teams prototype in Studio and migrate to ADK when they are ready to harden the agent for real workloads.

Step 4: Build and deploy

With your use case defined, build your first agent and deploy it through Agent Runtime. Keep this phase tightly scoped. The goal is one working agent in production, not a complete automation strategy.

Step 5: Govern and secure

Before expanding agent usage across the organization, governance needs to be properly configured. Set up Agent Identity so every agent has a trackable ID, configure Agent Gateway to control what each agent can connect to and what actions it can take, and enable Model Armor for protection against prompt injection.

Skipping this step before scaling is the single most common and costly mistake enterprises make with agent deployments.

Step 6: Optimize continuously

AI agents are not set-and-forget systems. They drift in production, edge cases emerge, and at scale, compute costs compound fast. A poorly scoped agent consuming unnecessary tokens can significantly inflate your bill without delivering proportional value.

Use Agent Evaluation to score agents against live traffic, Agent Observability to trace performance issues, and Agent Optimizer to automatically surface and fix failure patterns before they become expensive habits.

Get expert help deploying GEAP with revolgy

GEAP is comprehensive, and that comprehensiveness is what makes it complex to deploy well. Getting governance right, designing agents that perform in production, and connecting your existing systems securely requires experience most teams do not have. Revolgy is a certified Google Cloud Premier Partner that has done this before, across industries and at scale.

Revolgy deploys GEAP in three phases:

  • Phase 1: Foundation and Quick Wins: Revolgy runs discovery workshops to map your highest-value use cases, configures baseline security including Agent Identity, Gateway, and Model Armor from day one, and deploys quick-win agents that prove value internally before you scale.
  • Phase 2: Custom Agent Development: Revolgy builds purpose-specific agents using ADK, designed around your actual workflows and securely connected to your existing tools, including Jira, Salesforce, and other enterprise systems.
  • Phase 3: Adoption and Optimization: Revolgy handles team training, expands automation across the organization, and fine-tunes governance and cost controls to protect ROI as compute usage grows.

As an example, one of our clients was losing productivity to siloed information across Workspace, Jira, and Confluence. Revolgy deployed a Gemini Enterprise framework in four weeks with native connectors mapping user permissions across all platforms. Employees can now get answers synthesized from multiple systems simultaneously with zero compromise on data security.