Vertex AI becomes Gemini Enterprise Agent Platform: what changes for your company
Announced today at Google Cloud Next, Vertex AI has been absorbed into the Gemini Enterprise Agent Platform. Agent Runtime with sub-second cold start, persistent Memory Bank and code execution Sandbox arrive together. Zero disruption — but the roadmap is now different.
Fabiano Brito
CEO & Founder
What is the Gemini Enterprise Agent Platform
Agent Platform is not a brand new product — it is the unification of everything that existed in Vertex AI (model selection, Model Garden, Agent Builder, pipelines) with a new layer of agency tools: development, orchestration, DevOps, security and governance, all in a single interface.
⚡ Agent Runtime
Sub-second cold starts and new agent provisioning in seconds. Supports complex multi-step workflows and deep reasoning tasks requiring extended persistence.
- Cold start
- < 1 second
- Persistence
- Extended
🧠 Agent Memory Bank
Long-term memories curated from conversations. Memory Profiles offer high-accuracy recall with similarity search — context is never lost between sessions.
- Type
- Long-term
- Search
- Similarity
🔒 Agent Sandbox
Isolated and secure environment to execute model-generated code. Supports computer use tasks like browser automation — no risk to production environments.
- Isolation
- GKE Sandbox
- Computer use
- ✓ Browser
Vertex AI vs Agent Platform: what changed
| Capability | Vertex AI (before) | Agent Platform (now) |
|---|---|---|
| Model selection | Model Garden (200+ models) | Unified Model Garden + Gemini 3.1 Pro, Gemma 4, Claude |
| Agent development | Agent Builder (basic) | Agent Studio (low-code) + ADK (graph-based) |
| Execution | Standard serverless | Agent Runtime — sub-second, extended persistence |
| Memory | No cross-session memory | Memory Bank + Memory Profiles |
| Generated code | No native sandbox | Agent Sandbox — isolated and secure |
| Observability | Basic Cloud Logging | Agent Observability + Optimizer |
| Ready templates | — | Agent Garden: code modernization, financial analysis, invoice processing |
| Agentic security | — | Agent Gateway + Model Armor against prompt injection and data leakage |
| 3rd-party models | Via Model Garden | + Anthropic (Claude Opus/Sonnet/Haiku), MCP, AP2 |
The numbers Google disclosed
via ADK, Gemini models
Agent Runtime
Unified Model Garden
Customers already in production mentioned by Google: Color Health, Comcast, L’Oréal, PayPal, Burns & McDonnell, Geotab, Payhawk. These are agents running autonomously for days at a time, not just responding to one-off prompts.
What to do now — 4 practical steps
Your Vertex AI projects are already there. Go to console.cloud.google.com/agent-platform and explore the new Runtime, Memory Bank and Sandbox tabs.
Any support, service or process agent that today requires user "re-introduction" at each session is an immediate candidate for Memory Bank.
If your agent already generates and executes code (Python, SQL, scripts), moving to Sandbox eliminates security risks and simplifies compliance — without rewriting the logic.
Google provides pre-built agent templates — code modernization, financial analysis, economic research, invoice processing. Before building from scratch, check if an Agent Garden template solves 80% of your use case.
Vertex AI was the laboratory. Agent Platform is the factory. The difference isn't just a name — it's operational maturity.
Is your company ready for Agent Platform?
We help teams migrate Vertex AI workloads, implement Memory Bank in service agents and configure Sandbox for secure code execution — with auditable architecture and end-to-end support.