How do you run autonomous AI agents in production without losing control of security, compliance, or cost?
Broadcom has just announced VMware Tanzu Platform Agent Foundations a secure-by-default agentic runtime designed to take AI agents from experimental side-projects to governed, production-grade workloads running on VMware Cloud Foundation.
Unveiled at the AI in Finance Summit in New York, Tanzu Platform Agent Foundations extends the same code-to-production PaaS simplicity that Tanzu has always offered but now applied specifically to AI agents capable of autonomous decision-making.
Let’s Clear Up the Tanzu vs VKS Confusion
If you’ve been around VMware environments lately, you’ve probably heard both “Tanzu” and “VKS” thrown around, often interchangeably. They’re not the same thing. Here’s how they actually stack up.
VKS — vSphere Kubernetes Service is the Kubernetes engine itself. It’s what VMware built into vSphere to let you create and run Kubernetes clusters directly on top of your vCenter infrastructure. When you spin up a Kubernetes cluster on VCF, you’re using VKS to do it. Think of VKS as the engine in the car it’s doing the actual work of running containers, managing nodes, and handling scheduling. Your storage comes from vSAN, your networking from NSX-T, your control plane is managed by VMware instead of you bootstrapping it yourself like you would with k3s.
Tanzu Platform is the layer that sits on top of VKS. If VKS is the engine, Tanzu is the dashboard, the gear selector, and the sat nav all in one. Tanzu is what platform engineering teams use to define which applications can run, who can deploy what, which services developers are allowed to use, and how everything gets monitored and governed. Developers don’t interact with VKS directly but they interact with Tanzu Platform, and Tanzu talks to VKS underneath.
VMware Cloud Foundation (VCF) is the full stack underneath all of it — vSphere, vSAN, NSX-T, and all the infrastructure plumbing. VKS runs on VCF. Tanzu runs on VKS. They’re layers, not alternatives to each other.
So in plain terms: VCF is the datacentre, VKS is the Kubernetes engine inside it, and Tanzu is the platform that makes it usable for developers without them needing to understand all the infrastructure underneath.
┌─────────────────────────────────────┐
│ Tanzu Platform │ ← developers work here
│ (governance, apps, AI agents) │
├─────────────────────────────────────┤
│ VKS — vSphere Kubernetes Service │ ← Kubernetes engine
│ (clusters, nodes, scheduling) │
├─────────────────────────────────────┤
│ VMware Cloud Foundation (VCF) │ ← infrastructure
│ (vSphere, vSAN, NSX-T) │
└─────────────────────────────────────┘A Quick Word on MCP
MCP stands for Model Context Protocol.
An AI agent on its own is not very useful. It needs to connect to things a database, a ticketing system, a code repository, an API. Before MCP, every AI tool needed its own custom integration for each of those systems. MCP is the attempt to standardise that. MCP does the same thing for AI agents connecting to tools and data sources.
Broadcom’s announcement governs which MCP servers a developer’s AI agent is allowed to connect to, the same way you’d control which network a VM can reach. The platform team approves the list, developers work within it.
Why Agentic AI Needs a Different Runtime
Traditional applications execute instructions. AI agents also make decisions. That distinction changes everything from a security and operations standpoint. An agent with access to databases, APIs, and MCP servers is a significantly larger attack surface than a standard microservice. Highly regulated industries such as financial services, government, healthcare have been understandably cautious.
Broadcom’s answer is a runtime that bakes security in from the start rather than bolting it on. The platform uses Buildpacks instead of Dockerfiles for building agent containers, which means automatic patching and a consistent, auditable build process. Platform engineers curate what agents can access — models, MCP servers, marketplace services — before developers ever touch them. Same concept as how you’d control which datastores a VM can see in vCenter. Same idea, different layer.
What’s Actually in the Box
Tanzu Platform 10.4 ships with a concrete set of capabilities aimed at the full lifecycle of an AI agent — from standing it up to scaling it out.
Governed model access — IT curates pre-approved AI models, MCP servers, and marketplace services. Developers consume them; they don’t manage them. Same RBAC concept you’ll cover in your Kubernetes learning — who can access what is defined by the platform team, not left to individual developers.
Enterprise data engines — Tanzu for Postgres with pgvector, caching, streaming, and Spring AI memory services are all pre-integrated. Pgvector is an extension that lets Postgres store and search the kind of data AI models work with. From a storage perspective, it’s still Postgres on a PVC — just with an AI-specific extension on top.
Secure-by-default containers — Buildpack-based agent containers with automatic vulnerability patching. No Dockerfile sprawl, no manual patching cycles. Buildpacks are worth knowing — they’re an alternative to writing your own Dockerfile where the platform automatically figures out how to build your container image safely. Harbor, which is on your learning list, feeds directly into this.
VCF IaaS abstraction — Agents always get the compute, networking, and storage they need via programmatic VCF APIs. Tanzu is calling VCF APIs behind the scenes to provision storage and networking automatically, so developers never have to think about it. The infrastructure ownership still sits with you — the platform just automates the provisioning.
VKS-backed scalability — VMware vSphere Kubernetes Service underpins marketplace services, delivering elastic scale as agent workloads grow. This is the direct VKS connection — everything in Tanzu Platform that needs to scale runs on VKS underneath.
Pre-built agent quickstart — Developers can start from a pre-built agent template, accelerating the move from proof-of-concept to production deployment.
The Storage and Infrastructure Angle
For infrastructure practitioners, the most relevant part of this announcement is how Tanzu Platform abstracts VCF’s IaaS APIs to ensure agents always have the resources they need. That means storage provisioning, network policy, and compute scheduling are handled programmatically — platform teams define the guardrails, VCF enforces them, and developers never have to file a ticket to get a persistent volume.
The integration of Tanzu for Postgres with pgvector is notable for anyone building retrieval-augmented generation (RAG) workloads. Instead of standing up a separate vector database, teams can use a managed, enterprise-supported Postgres instance within the Tanzu ecosystem — with all the backup, HA, and lifecycle management that comes with Tanzu Data Services. The interesting question from a storage angle is how vSAN backs these Postgres instances — and the answer is the same StorageClass and PVC model on k3s, just with vSAN as the backend instead of local-path.
How This Fits into the Broader VCF 9 Story
This announcement doesn’t exist in isolation. Broadcom has been steadily making VCF 9 “AI native” — with Private AI Services included in the standard subscription at no additional cost. Today’s Agent Foundations announcement completes a key piece of that picture: it’s not enough to run models privately; you also need a runtime that lets those models act autonomously, safely, and at enterprise scale.
The MCP governance story is worth watching closely. Tanzu Platform’s ability to centralise and govern MCP server integrations — with observability into usage patterns so teams can tune, optimise, or retire servers — positions VCF as a serious control plane for enterprise agentic architectures.