The stack: umbrella, framework, runtime

Lesson 1 of 5 in Google’s Agent Stack: ADK, Agent Runtime, and the Names That Keep Moving.

Current as of September 2026 — and Google’s names move faster than this page

Read this module for the shapes: what each piece of Google’s agent stack does, what it costs you in coupling, and what questions to ask. Do not read it as vendor documentation, because in this corner of Google Cloud the labels have changed at least three times in two years.

What was checked in September 2026: the product page is titled “Gemini Enterprise Agent Platform (formerly Vertex AI)”; the Vertex AI documentation carries a banner saying it is no longer being updated; the managed runtime documented for a year as Vertex AI Agent Engine is presented as Agent Runtime; and the ADK docs at google.github.io/adk-docs now 301-redirect to adk.dev.

What was not established: the date Google announced the rebrand, the release date of ADK 2.0, and the launch stage (GA vs preview) of most individual platform features — the overview page does not print them. So treat every product name below as “what the docs said in September 2026”, and check the current docs before you put a name in a design document.

Google’s stack is easiest to hold if you split it in half along a single line: the framework you write code in versus the platform that runs it.

On the framework side sits ADK — the Agent Development Kit — which describes itself as “the open-source agent development framework that lets you build, debug, and deploy reliable AI agents at enterprise scale.” It is code-first, it is open source, and it runs perfectly well on your laptop or in a container you own. Nothing about ADK requires Google Cloud.

On the platform side sits the managed runtime plus a ring of managed services around it: Sessions, Memory Bank, Example Store, evaluation, a code-execution Sandbox, a Feedback service, Agent Gateway, and a Skill Registry. That side does require Google Cloud, bills by the hour, and is where the branding churn concentrates.

The umbrella over both halves is the part that keeps being renamed. “Vertex AI Agent Builder” was the 2025 name for the suite — ADK, the managed runtime, Agent Garden, and the tooling around them — and in September 2026 that branding is folded into Gemini Enterprise Agent Platform, the successor umbrella to Vertex AI itself. A Google Cloud editor’s note puts it plainly: “Vertex AI has evolved. We’ve launched Gemini Enterprise Agent Platform, the new home for building, scaling, governing, and optimizing your agentic workforce.”

Confusingly, “Agent Builder” has meant three different things. Say the capability, not the brand: “the open-source framework”, “the managed runtime”, “the memory service”. Capabilities survive rebrands; product names do not.

“Agent Builder” — which one do you mean?

Three distinct referents, all real:

The 2024 product — Google’s search-and-conversation builder. On 9 April 2025 it was renamed AI Applications, with functionality and endpoints unchanged.
The 2025 suite — on the same day, “Vertex AI Agent Builder” was redefined to mean a suite of features for building and deploying agents in Vertex AI: ADK, the managed runtime, Agent Garden.
The 2026 umbrella — that suite branding now appears under Gemini Enterprise Agent Platform, per the product page title checked in September 2026.

If a blog post, a Stack Overflow answer, or a colleague says “Agent Builder”, ask which year they mean before you believe the feature list.

Agent Engine or Agent Runtime?

Same lineage, three names. LangChain on Vertex AI (the Reasoning Engine era) was renamed Vertex AI Agent Engine, which reached general availability on 4 March 2025 with billing starting the same day. As of September 2026 the docs present it as Agent Runtime inside the Agent Platform, described as “a fully managed environment for developers to handle testing, release management, and reliability at a global scale” where you “deploy and scale agents efficiently without the need to manage underlying infrastructure.” There is also an “Agent Runtime SDK migration” guide, which tells you the SDK surface moved too.

Practical consequence: search results, tutorials, and half of Google’s own corpus still say Agent Engine. When you name it in your own documents, name both.

And the Gemini Enterprise app — is that the same thing?

No, and this is the distinction people get wrong most often. Gemini Enterprise (the app) launched 10 October 2025 as “the new front door for AI in the workplace” — intranet search, an AI assistant, and an agentic platform for knowledge workers, with permissions-aware enterprise search, prebuilt connectors (Confluence, Jira, Microsoft SharePoint, ServiceNow), an Agent Gallery, and editions named Business, Standard, Plus, Pay-as-you-go, and Frontline.

That is an end-user product. The Agent Platform is the developer platform. They share a brand family and a governance story; they are not interchangeable. The app is positioned as the successor to Google Agentspace — its docs still live under /agentspace/ URL paths — though no page checked in September 2026 states an explicit rename, so describe it as succession, not renaming.

What got retired on the way here?

Vertex AI Extensions — the earlier way to give a Gemini model external capabilities — was deprecated on 26 May 2026 and shuts down after 26 November 2026, with Google recommending migration to Agent Platform. That is the concrete cost of building on a pre-1.0 vendor abstraction: not a rename you can ignore, a migration with a date on it.

Key terms: agent, agent loop, tool, memory, MCP, A2A

Build → deploy → operate: which piece owns which phase (Agent Platform docs and adk.dev, checked September 2026)
Phase	What you are deciding	Google’s piece	Can you leave it behind?
Build — write the agent	Agent structure, tools, models, prompts, orchestration between sub-agents.	ADK, in Python, TypeScript, Go, Java, or Kotlin. Open source, model-agnostic, runs anywhere.	Yes — cheaply. It is a library in your repo. This is the least sticky layer of the stack.
Deploy — get it running	Where the process lives, how it scales, how revisions roll out, who can reach it.	Agent Runtime (formerly Agent Engine) — or Cloud Run, GKE, or a container you host yourself. ADK documents all four targets.	Yes, if you keep the agent’s state and telemetry portable. The runtime is a hosting choice, not a rewrite.
Remember — state across turns and sessions	Short-term conversational state; long-term facts and preferences per user.	Sessions (individual interactions between users and agents) and Memory Bank (long-term memories generated from sessions).	Hardest to leave. The memory record shape and the extraction behaviour are the sticky parts, not the API call.
Reach tools — call the outside world	Protocol conversion, credentials, discovery, authorization on each call.	ADK tool types (function, MCP, OpenAPI, built-ins) plus Agent Gateway at platform level and a Skill Registry for storing and dynamically loading skills.	Partly. Tool definitions travel with your code; brokered credentials and gateway policy do not.
Operate — run it in production	Traces, logs, revisions and traffic splits, private networking, provisioning as code.	Documented runtime operations include Cloud Trace, Cloud Logging, monitoring, revision and traffic management, bidirectional streaming, Private Service Connect interfaces, Terraform provisioning, and IAM-based agent identity.	Only if you emit your own traces as well. Trace format is a quiet lock-in vector on every cloud.
Govern — decide what it may do	Content filtering, action-level policy, encryption, residency, audit.	Gemini safety filters, Semantic Governance Policies over tool calls, plus VPC Service Controls, CMEK, and data residency at rest per the platform’s enterprise-security matrix.	No — and you should not want to. Enforcement belongs next to execution, wherever that is.

One ADK agent, three plausible destinations

ADK agent in your repo
Python, TypeScript, Go, Java, or Kotlin. Same agent definition regardless of where it ends up running.
Run and debug locally
ADK is a library first: the loop runs on your machine with no Google Cloud dependency at all.
Who should operate it?
This is the real fork. Not “which framework” — the framework is already chosen — but how much operational surface you want to own.
Agent Runtime (managed)
Managed infrastructure, built-in authentication, Cloud Trace observability, revisions and traffic management. Metered by allocated vCPU and RAM.
Cloud Run or GKE
You keep the container contract and your own scaling, networking, and telemetry choices. ADK documents both as deployment targets.
Self-hosted container anywhere
Any cloud, any orchestrator, your own secrets and traces. Cheapest exit, largest backlog.
Sessions, Memory Bank, Gateway — opt in per service
The managed services are separable. Self-hosting the loop does not preclude buying the memory service, and vice versa.

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.