Broadcom Has Built Its Answer to Platform Engineering 2.0

In July, I asked what the Platform Engineering 2.0 platform is actually made of. At VMware Explore, Broadcom answered with its own bill of materials.

That answer did not arrive as a single product. It came through announcements spanning VMware Cloud Foundation, VMware AI Factory, AI-ready data foundations in Tanzu Platform, AgentMinder, TrueSource, vDefend and Avi Load Balancer. Viewed separately, each is another addition to an already large portfolio. Viewed together, they look much more like a productized reference architecture for Platform Engineering 2.0.

Mike Vizard has already reported the details of the VMware AI Factory and TrueSource announcements. I am not going to replay them feature by feature. The more important question for platform teams is what the combined portfolio says about where platform engineering is headed.

Broadcom has now connected infrastructure, models, data, agents, identity, software artifacts, security and cost controls into one architectural proposition. It is an ambitious answer. It is also one that raises important questions about availability, overlap, openness and whether a tightly integrated platform can remain genuinely composable.

One Platform Story Hidden Inside Several Announcements

The working model I laid out in July began with an experience layer. Beneath that sat orchestration and delivery, a workload control plane, an AI and agent control plane, virtualized infrastructure and the physical compute, accelerators, storage and networking on which everything ultimately depends. Security, identity, observability, governance and FinOps needed to operate through the entire stack rather than sit alongside it.

Broadcom’s VMware Explore portfolio maps surprisingly well against that model.

At the foundation, VMware AI Factory combines VMware Cloud Foundation with hardware, accelerators, model-serving software and infrastructure automation. MetalSoft adds heterogeneous bare-metal provisioning. VMware Cloud Foundation pools compute, storage, networking and GPU resources while providing the isolation and operational control needed to share expensive infrastructure.

Above that, Tanzu provides the application and agent delivery environment. Its new agent foundations include deny-by-default sandboxes, isolated credential management, approved tools, an agent harness, persistent memory, human-review controls and a governed marketplace for models, services, MCP servers, data products and agents.

AgentMinder provides another control layer. Broadcom says it verifies agent identity and evaluates individual actions against the agent’s declared mission, intent, context and current risk. VMware vDefend and Avi extend network and runtime protections to models, agents, MCP servers, tools and sensitive data.

TrueSource addresses what applications and agents are built from, supplying verified open-source artifacts, hardened images, supported dependencies and data services. Broadcom’s AI-ready data foundations address what agents know by turning fragmented enterprise information into governed data products with context, access controls and lineage.

The result is an attempt to connect metal to model, model to agent and agent to enterprise action. Network World framed the announcements primarily as Broadcom’s bet on private AI. That is certainly part of it. From a platform engineering perspective, however, Broadcom is doing something broader: It is proposing an operating system for how enterprises build, run and govern both human- and machine-produced work.

The Platform Has a New Kind of Customer

Platform Engineering 1.0 primarily mediated between developers and infrastructure. Developers followed a golden path, and the platform fulfilled their requests for environments, pipelines, services and deployment capabilities.

Platform Engineering 2.0 must also mediate between autonomous software and the enterprise.

As Nathan Eddy recently explored in “When Your Next Platform User Is an AI Agent,” agents do not primarily navigate graphical portals. They consume APIs, request resources, call tools, retrieve data and sometimes delegate work to other agents. They may provision infrastructure, open pull requests or attempt remediation without waiting for a human to perform each step.

That requires machine-consumable versions of capabilities platform teams previously designed for people: Identity, entitlements, golden paths, resource limits, cost constraints, approved tools, policy decisions and evidence of what happened.

It also introduces a different risk model. An agent can make the wrong request repeatedly, consume resources nonlinearly or search for another path after encountering a denial. A platform designed for human-paced consumption cannot simply give an agent a service account and hope the existing guardrails hold.

Platform Engineering 1.0 fulfilled developer requests. Platform Engineering 2.0 must decide whether machine intent should be allowed to become enterprise action.

The Control Plane Becomes the Strategic Product

That makes AgentMinder one of the more consequential parts of Broadcom’s portfolio.

Authentication answers who or what is making a request. It does not answer whether this particular action is appropriate, whether it fits the agent’s assigned mission or whether the underlying risk has changed since permission was originally granted.

Broadcom says AgentMinder evaluates those conditions before an action reaches a model, API, MCP server, tool or enterprise resource. The platform can allow, deny, redirect or redact the request while recording the resulting decision.

The use of OpenID AuthZEN is notable. The OpenID Foundation’s work on authorization for the agent era addresses a question that ordinary OAuth scopes cannot adequately express: May this agent, acting for this user, call this tool with these arguments under the present conditions?

That granularity becomes more important as agents delegate work. PlatformEngineering.com recently examined how least privilege can deteriorate across an agent chain. A credential may be appropriately scoped when first issued, but nothing automatically guarantees that authority becomes narrower as the request passes through additional agents and tools.

OpenTelemetry provides the evidence layer in Broadcom’s model. It can record prompts, calls, policy decisions and agent actions in a format that observability systems can consume. That is necessary, but it is not sufficient. A complete trace can tell us exactly what an agent did without telling us whether its conclusion was correct, useful or worth the resources consumed.

Observability can establish that the system operated as designed. Evaluation must establish whether the design produced the desired outcome.

AI Makes Infrastructure Visible Again

In my previous article, I argued that the hypervisor never disappeared. Public clouds did not eliminate virtualization. They absorbed it and made the underlying machinery less visible to customers.

AI is pulling back that curtain.

GPUs are too expensive, scarce and operationally demanding to treat as unlimited commodities. Models have different memory, compute and latency requirements. Agentic workflows can generate variable infrastructure and token consumption, including when the work ultimately fails. Sovereignty, privacy and data-residency rules may determine where inference is allowed to occur.

VMware AI Factory is Broadcom’s infrastructure answer. It provides shared and isolated GPU pools, model serving, multitenant access, hardware and model choice, and visibility into compute and token consumption. Broadcom says automation can reduce the path from bare metal to the first functioning model from weeks to hours. That claim will need to be validated across customer environments, but the architectural direction is clear.

Infrastructure has again become an explicit part of the platform product.

That does not mean every enterprise should move AI workloads into a private cloud. An organization consuming models primarily through hosted APIs will have different needs from one operating its own inference infrastructure. Many will do both. Others will choose public cloud, colocation or bare metal for specific workloads.

Broadcom’s strongest proposition is evolutionary. Enterprises with substantial VMware estates, mixed virtual machine and Kubernetes workloads, regulated data or a need for greater infrastructure control can extend what they already operate rather than create a completely separate AI island.

Data Belongs Inside the Platform

Broadcom’s AI-ready data foundations may ultimately prove as important as the infrastructure announcements.

Agents are only as reliable as the information available to them. Enterprise data is rarely clean, complete and neatly assembled for model consumption. It is duplicated, contradictory, partially migrated, poorly structured and spread across documents, databases and object stores.

Broadcom is proposing a data-to-AI lifecycle layer that can ingest structured, unstructured and multimodal information, add semantic context and lineage, and publish the result as governed data products. An embedded MCP server can then make those products available to approved agents without giving those agents unrestricted access to raw backend systems.

Fresh Signal65 PINNACLE results provide independent evidence for why this belongs in the platform architecture. Forty-three of 44 tested model configurations lost ground when they moved from organized data to data as enterprises actually maintain it. The median configuration surrendered approximately 28 points of workflow completion between the two conditions.

A better model does not eliminate a broken information environment. Nor does connecting a capable model to an MCP server automatically create a reliable enterprise agent.

This suggests that data and knowledge authority deserve treatment as a first-class platform concern. Whether we model it as a separate plane or a capability operating across the stack, the platform must govern not only which data an agent can reach, but which version is authoritative, how its lineage is established and how context is kept current.

The platform cannot promise reliable agents merely by supplying intelligence. It must govern the context on which that intelligence acts.

Trust Starts Before Runtime

TrueSource belongs in this architecture for a similar reason.

AI is accelerating code production. It is also accelerating vulnerability discovery and potentially shortening the time between disclosure and exploitation. Platform teams therefore need trusted inputs, provenance and validated remediation incorporated into their paved roads.

TrueSource combines supported Spring software with clean-room builds, signed artifacts, hardened Bitnami images and supported open-source data services. Broadcom is using AI to help scan code and dependencies, but it is retaining human verification before distributing patches.

That is an important boundary. AI can accelerate the engineers responsible for software supply-chain security without becoming the final authority on what is safe to deploy.

Platform Engineering 2.0 must govern more than where a workload runs and what it can access. It must also provide assurance about what the workload is made from.

Integration Is Not the Same as Composability

There is a great deal to like in Broadcom’s architectural direction. Agents become first-class platform users. Identity is connected to action-level authorization. Credentials remain outside agent sandboxes. Data products become governed resources. Infrastructure utilization is connected to token economics. Existing investments can be extended instead of discarded.

But the announcements also leave questions that enterprises should ask before treating the architecture as complete.

AgentMinder and TrueSource are available now. Several Tanzu capabilities are expected in fall 2026, while elements of the vDefend, Avi and AI Factory roadmaps remain forthcoming. It would be premature to discuss every component as though it were already shipping and integrated in customer production environments.

Broadcom also now describes gateway and control functions across VCF, Tanzu and AgentMinder. Platform teams will need clear boundaries among them. Which layer owns identity? Which enforces tool policy? Which tracks token consumption? Which supplies the authoritative audit trail?

Then there is composability. Broadcom supports multiple models, accelerators and hardware partners. That provides choice within the architecture, but it does not by itself demonstrate that the architecture is composable. How easily can a customer replace its model server, agent framework, data service, gateway or observability platform? Can AgentMinder govern agents running outside a predominantly VMware environment as effectively as it governs workloads inside one?

An opinionated platform should reduce complexity by giving teams a paved road. The risk is that integration becomes dependency and the paved road becomes another golden cage.

Broadcom also reports more than 3,000 VCF 9 deployments, but as Mike Vizard noted, the percentage running production AI workloads remains unknown. The architecture is substantial. The adoption evidence for the complete AI and agent stack still has to accumulate.

From Vendor Architecture to Practitioner Roadmap

Broadcom has now supplied one coherent answer to what Platform Engineering 2.0 could look like. It will not be the only answer.

PlatformEngineering.com and Futurum Research are examining what practitioners are actually building. Are enterprises extending their existing developer platforms or creating separate AI platforms? Are they using hosted models, private infrastructure or hybrid approaches? Who owns agent governance? Where does data and knowledge authority reside? Which capabilities are being built, bought or consumed as services?

We also want to identify whether recognizable platform archetypes are emerging and what enterprises should prioritize during the next 12 to 18 months. Platform practitioners can help shape that work by completing our short Platform Engineering 2.0 survey.

Platform Engineering 2.0 began as an argument that platforms designed for human developers and conventional applications would need to expand. Broadcom has now placed products beneath that argument.

Its VMware Explore portfolio does not define a universal Platform Engineering 2.0 stack. It does give enterprises a concrete architecture to evaluate, and it gives the rest of the platform market an answer it must address.

The competition will not be decided by which vendor assembles the longest feature list. It will be decided by which platforms give humans and agents an approved way to build, run, observe, secure, govern and pay for applications, models, data and autonomous work while avoiding another generation of infrastructure lock-in.