Google Gemini Enterprise Agent Platform — Vertex AI gets a real agent stack
At Google Cloud Next on April 22, 2026, Google introduced the Gemini Enterprise Agent Platform — and made it GA on the spot. What used to be Vertex AI Agent Builder is now part of a new, integrated platform that pulls model selection, agent building, runtime, memory, and governance under one roof. The most important structural detail: “All Vertex AI services and roadmap evolutions will be delivered exclusively through the Agent Platform.” Vertex AI as a standalone product is going away.
What the platform actually bundles
- Agent Studio — low-code visual builder for agents.
- Agent Development Kit (ADK) — code-first environment; according to Google it processes 6+ trillion tokens per month.
- Agent Runtime (formerly “Agent Engine”) — sub-second cold starts, support for multi-day workflows.
- Memory Bank (Sessions are now “Agent Platform Sessions”) — persistent long-term context, $0.25 per 1,000 events or memories since January 28, 2026.
- Agent Registry and Agent Gateway — central tool library, “air traffic control” for agents with security policies.
- Model Garden — 200+ models including Gemini 3.1 Pro, Gemma 4, Anthropic Claude.
What used to be true
Vertex AI was a grab-bag: Model Garden, Agent Builder, Vector Search, MLOps tools — solid building blocks, loosely connected. Anyone running agents in production juggled multiple consoles, pricing sheets, and lifecycle patterns. An agent that needed to hold state over multiple days was a custom build using Cloud Tasks, Firestore, and bespoke glue. Memory across sessions: your own data model.
Google itself doesn’t pitch the old world as a weakness, but the new claim is direct: “We’ve simplified the entire lifecycle with AI-native coding capabilities to help you ship production-grade agents faster.”
What’s true now
1. One platform, one lifecycle. Agent Studio is the entry point — prompt in, agent out. If that’s not enough, you export to the ADK and continue code-first without losing context. That’s the real change: low-code and code-first are no longer two separate worlds with a translation tax — they’re a continuous path.
2. Multi-day workflows as a default. The new Agent Runtime holds state across days without you having to manage persistence, resume logic, or cold-start handling. That opens up use cases that were finicky with classic function-as-a-service patterns — long-running research jobs, onboarding flows with wait states, agents that wait on external events.
3. Memory Bank as persistent context. Instead of manually managing conversation history per session, the agent writes “memories” into the Memory Bank — structured long-term records that persist across sessions. Pricing since January 28, 2026: $0.25 per 1,000 events or memories — cheap enough that memory can be on by default.
4. Agent Identity for governance. Every agent gets a unique cryptographic ID. That makes agent actions auditable in the same way service-account actions are. For regulated industries this is the first clean audit model for autonomous agents.
5. Agent Garden with templates. Pre-built templates for code modernization, financial analysis, invoice processing, and other common use cases. Combined with Native Ecosystem Integrations (plug-and-play connectors for Workspace, BigQuery, Salesforce, etc.), the path from “hello agent” to a productive workflow is now hours, not days.
Where the marketing holds up — and where it doesn’t
The bundling is real and useful. Lifecycle breaks between builder, runtime, and memory were the single most common bottleneck in Vertex AI agent projects — those are gone. Sub-second cold starts and multi-day workflows are concrete technical wins that older Vertex setups couldn’t match.
The downside: bundling deepens lock-in. Anyone using Agent Studio writes workflow logic in a proprietary format that doesn’t port trivially to AWS, Azure, or on-prem. Memory Bank and Agent Identity are equally Google-specific. If you have a multi-cloud strategy, you’ll weigh that against the convenience.
Second observation: Google opens Model Garden with 200+ models including Anthropic Claude — that’s an admission that Gemini alone doesn’t cover the spectrum. For customers who want model diversity, that’s a clear plus; for Gemini’s strategic position, it’s a hint that Google isn’t clearly winning the model race.
What you can do now
If you sit on Vertex AI: check the migration path over the coming weeks. Google is sunsetting Vertex AI as a standalone — all new features land exclusively on the Agent Platform. Existing Vertex workloads keep running, but roadmap updates won’t appear there anymore.
If you’re starting fresh on Google Cloud: begin directly with Agent Studio for first iterations and export to ADK as you grow. The path is cleaner than the old combination of Vertex AI Agent Builder plus manual MLOps.
If you’re planning multi-cloud: the platform is powerful but heavily lock-in. Consider keeping workflow logic in a portable format (Agents SDK, LangGraph, your own routing) and using platform features selectively — Memory Bank, for example, is usable standalone.
Entdecke mehr
Why AI Models Find Different Code Problems
Three frontier models, the same 1000-line script, three different finding lists — and why that very spread makes multi-orchestration strong.
GlossarEnsemble / Multi-Model Orchestration
Ensemble means combining several deliberately varied LLM runs or models whose findings complement each other. Multi-model orchestration drives these runs via orchestrators with sub-agents, so the union of results is larger than any single run.
LexikonFunction Calling / Tool Use
How an LLM uses tools: define a tool as a schema, the model picks the function and arguments, the result returns to the chat — the basis of every agent.