Gemini Enterprise Agent Platform (GEAP) is Google’s April 2026 platform for the full agent lifecycle — “build, scale, govern, and optimize agents” in Google’s own framing. It’s the umbrella that now sits over what used to be three separate stories: Vertex AI (see What is Vertex AI) as the ML platform, Agentspace (launched December 2024) as the enterprise agent product, and What is Google ADK as the code-first build framework. Agentspace was rebranded Gemini Enterprise roughly ten months after launch; this platform is the next consolidation on top of that.
The four-stage lifecycle
Everything else in this note is a component slotted into one of four named stages:
Build Scale Govern Optimize
"choose the right "clear the path "establish "guarantee
environment" to production" centralized control" quality"
│ │ │ │
Agent Studio Agent Runtime Agent Identity Agent Simulation
ADK Memory Bank Agent Registry Agent Evaluation
Agent Garden Agent Sandbox Agent Gateway Agent Observability
Semantic Gov Policies Agent Optimizer
Build
| Component | What it is |
|---|---|
| Agent Studio | Low-code visual canvas for designing agent reasoning loops |
| What is Google ADK | Code-first, model-agnostic framework — “AI-native coding capabilities” |
| Agent Garden | Curated templates (code modernization, financial analysis, invoice processing) |
This is the same Agent Builder split already covered in What is Vertex AI — GEAP is the branding it now ships under.
Scale — a runtime built for long-running agents
The Agent Runtime was explicitly re-engineered to support agents that “maintain state for days at a time,” not just request/response calls. Two things make that possible:
- Memory Bank — persistent, long-term context, functionally the same problem What is Mem0 solves standalone, except scoped to this runtime.
- Agent Sandbox — a hardened execution environment for model-generated code, so a long-running agent can safely run code it wrote itself rather than only calling pre-registered tools.
The metered vCPU/memory pricing for this runtime layer (formerly “Agent Engine”) is detailed in What is Vertex AI rather than duplicated here.
Govern — the control plane
This is the part of GEAP that doesn’t have an equivalent in What is CrewAI or a self-hosted What is Hermes Agent setup — it’s the reason an enterprise buys the managed platform instead of assembling the pieces:
| Component | Role |
|---|---|
| Agent Identity | Unique cryptographic ID per agent — the basis for access control and audit |
| Agent Registry | Single source of truth indexing every internal agent, tool, and skill — including registered MCP servers |
| Agent Gateway | ”Air traffic control” — central policy enforcement for all MCP + A2A traffic |
| Semantic Governance Policies | Natural-language constraints applied to an agent’s tool calls, not just RBAC on the tool itself |
Model access and discovery
- Model Garden — first-class access to 200+ models, not just Google’s own: Gemini 3.1 Pro, Claude Opus/Sonnet/Haiku, Gemma 4, Lyria 3. Anthropic is a named first-party partner here, the same multi-model posture What is Google ADK has at the framework level.
- Agent Gallery — the discovery surface for agents, not models: validated agents from partners including Adobe, Salesforce, ServiceNow, and Workday, browsable inside the Gemini Enterprise app.
Optimize — closing the loop
| Component | What it does |
|---|---|
| Agent Simulation | Tests agents against synthetic, human-like user interactions |
| Agent Evaluation | Continuously scores agents against live production traffic |
| Agent Observability | Full execution traces — a real-time view into agent reasoning, not just outcomes |
| Agent Optimizer | Clusters real-world failures automatically and suggests refined system instructions |
This is a superset of the evaluation harness What is Google ADK ships (criteria-based scoring, user/ environment simulation) — GEAP adds the always-on observability and auto-clustering of failures on top.
Interop: A2A + MCP as first-class traffic, not an afterthought
The Agent2Agent (A2A) protocol is Google’s open protocol so agents from different builders and platforms can discover each other, collaborate, and securely delegate tasks — the piece that lets a partner agent from Agent Gallery talk to an in-house ADK agent without custom glue code. Agent Gateway enforces policy across both A2A and MCP traffic uniformly, and Agent Registry treats MCP servers as first-class citizens alongside agents and skills. Real deployments already lean on this: L’Oréal’s Beauty Tech Agentic Platform pairs ADK with MCP; PayPal layers its own Agent Payment Protocol (AP2) on top for multi-agent commerce workflows.
Pricing shape (2026)
Platform-level meters, distinct from the Agent Runtime’s own vCPU/memory billing (see What is Vertex AI):
| Meter | Rate |
|---|---|
| Agent Storage | $0.30 / GiB-month |
| Agent Compute | $0.085 per 3M read operations |
Editions are named Standard, Plus, Frontline, and Business — the public pricing page splits entitlements across these but doesn’t fully enumerate what’s gated per tier as of this writing, so treat the edition boundary as directional until confirmed against a live quote.
Where it fits
| Concern | This note (GEAP) | What is Google ADK | What is Vertex AI |
|---|---|---|---|
| Scope | Umbrella platform: build+scale+govern+optimize | Code-first build framework only | Deploy runtime + historical ML platform |
| Governance (Identity/Registry/Gateway) | Native, platform-level | None — bring your own | Not covered |
| Agent discovery | Agent Gallery (partner agents) | N/A | N/A |
| Multi-agent interop | A2A protocol, native | A2A client support | N/A |
Why it’s on the backlog: the governance stack (Agent Identity, Agent Registry, Agent Gateway) is the piece that would actually matter if the h-aiops SRE-agent line ever needs to run multiple agents with different trust levels against ShipSolid’s production tooling — audit and policy enforcement per-agent, not just per-service-account. The A2A protocol is the other reason to watch this: it’s what would let a home-grown ADK agent interoperate with a partner-built agent from Agent Gallery without custom integration work.
Local graph
Linked from 4 notes
What is Google ADK
Google's open-source, code-first Agent Development Kit — a multi-language framework for building, evaluating, and deploying agents, positioned as an 'agent execution framework' rather than a toolkit.
What is Vertex AI
Google Cloud's managed ML/AI platform — as of 2026 rebranded and consolidated into the Gemini Enterprise Agent Platform, bundling 200+ foundation models, Agent Builder, and a managed agent runtime (formerly 'Agent Engine').
Agentic AI: Projects & Engineering Mastery
A book-shaped table of contents for Agentic AI: Projects & Engineering Mastery: hands-on practitioner builds, Principal/Staff-level technical leadership, and the lookup appendices and vendor/framework reference notes for the whole series. Book 6 of the AI Systems Engineering series.
10. Agent Meshes
Covers service-mesh-inspired architectures for agent-to-agent discovery, routing, and observability at the scale of dozens of interacting agents.
Related notes
What is Google ADK
Google's open-source, code-first Agent Development Kit — a multi-language framework for building, evaluating, and deploying agents, positioned as an 'agent execution framework' rather than a toolkit.
What is Vertex AI
Google Cloud's managed ML/AI platform — as of 2026 rebranded and consolidated into the Gemini Enterprise Agent Platform, bundling 200+ foundation models, Agent Builder, and a managed agent runtime (formerly 'Agent Engine').
What is CrewAI
Open-source Python framework for multi-agent orchestration — role-based 'Crews' for autonomous collaboration and event-driven 'Flows' for precise control, now a de facto standard for production agentic pipelines.
What is Hermes Agent
Nous Research's open-source, self-hosted AI agent that runs continuously on your own server, builds persistent memory over time, and distills its own reusable 'skills' from completed tasks.