The 40% Threshold
Gartner's projection that 40% of enterprise applications will embed AI agents by 2026 has been cited so often it risks becoming background noise. But pair that figure with the CSA survey finding that 78% of organizations lack agent identity policies, and a different picture emerges. We are not witnessing gradual adoption. We are witnessing a governance vacuum forming in real time, one where the tools we built to manage software identities simply do not extend to autonomous agents making API calls on behalf of enterprises.
Over the past seven days, I have been tracing the infrastructure implications of Broadcom's AgentMinder announcement at VMware Explore 2026. The product itself is straightforward—a central control plane for AI agent identity, permissions, and traffic routing. But beneath the press release lies a structural argument about where AI governance belongs in the enterprise stack, and that argument deserves closer scrutiny.
The Service Mesh Precedent
AgentMinder's architecture should look familiar to anyone who spent the late 2010s wrestling with microservices. The "central control plane and traffic controller" positioning is essentially service mesh thinking applied to AI agents—the same pattern Istio and Linkerd brought to service-to-service communication, now adapted for agent-to-API interactions. Broadcom has taken the policy-as-code and API gateway patterns that matured in the microservices era and retrofitted them for autonomous agents.
This explains the unusually high technical maturity for a product category that barely existed eighteen months ago. Broadcom's internal deployment processes 43 million API calls daily across 20 million customer identities and 72,000 employee identities, with the CIO claiming zero downtime at global scale. These are production numbers, not pilot metrics. The company has effectively dogfooded its own governance solution at a scale most enterprises will not approach for years.
The integration choices are telling. By building on AuthZEN for authorization and OpenTelemetry for observability, Broadcom signals an intent to avoid proprietary lock-in at the protocol level—even as the deployment itself remains deeply coupled to the VMware stack. That tension between open standards and closed infrastructure defines AgentMinder's strategic position.
The Infrastructure Layer Argument
Broadcom's bet is that governance belongs in the infrastructure layer, not the identity layer. The logic is straightforward: if you control the traffic path, you can enforce policy in real time rather than relying on事后 audits. An agent can only route to authorized backends, can only invoke approved tools, can only act within task-bound permissions. This is "forced safety"—external constraints that prevent unauthorized action regardless of what the model might otherwise attempt.
For compliance-driven industries—financial services, healthcare, government—this has obvious appeal. The chain of custody features and audit-grade visibility speak directly to regulatory requirements that identity-layer solutions struggle to meet. When a regulator asks what an agent did, infrastructure-layer governance can produce a definitive answer from the traffic path itself.
But the approach carries structural costs. The "zero downtime" claim implies multi-region active-active deployment with hot upgrade capabilities. The 43 million daily API calls require dedicated gateway infrastructure with meaningful performance headroom. This is not a lightweight SaaS deployment; it is a significant infrastructure commitment that presupposes an existing VMware investment.
The Blind Spot in the Control Plane
Here is where the analysis gets uncomfortable. AgentMinder addresses what agents can do—permission boundaries, traffic routing, token authentication. It is far weaker on what agents should do and what agents actually did. Prompt injection attacks operate at the content layer, invisible to traffic-level governance. Agent-to-agent communication, arguably the most dangerous attack surface in multi-agent systems, receives no explicit coverage in the announced feature set. And the central control plane itself becomes a high-value target: compromise the governance layer, and you compromise every agent under its authority.
Based on my experience auditing cross-chain bridges during the 2022 crisis, I recognize this pattern. Centralized control points solve coordination problems but introduce single points of failure. The bridges that survived the Terra collapse were those with distributed liquidity reserves and redundant validation paths. The ones that failed had concentrated authority that became the attack surface. AgentMinder's architecture concentrates governance authority by design, which is precisely what makes it effective—and precisely what makes it vulnerable.
The "zero downtime" requirement also creates an inherent tension with security. High-availability architectures often require bypass paths and degraded modes that can be exploited. The question Broadcom has not yet answered publicly is how the control plane protects itself, not just how it protects the agents beneath it.
The Okta Counterfactual
The competitive framing against Okta's Agent SSO reveals more than product positioning. Okta represents the identity-layer thesis: governance as a platform-agnostic service that works across any infrastructure, deployable in days rather than quarters. Broadcom represents the infrastructure-layer thesis: deeper enforcement at the cost of platform commitment.
Both approaches have merit, but they are not symmetric. Identity-layer governance scales across heterogeneous environments and captures the long tail of agent deployments. Infrastructure-layer governance provides stronger enforcement for the critical systems that justify the implementation cost. The market will likely segment accordingly—comprehensive coverage from identity providers, deep enforcement from infrastructure vendors.
The hidden variable is the cloud providers. AWS, Azure, and GCP are all building agent governance capabilities into their platforms. If Amazon Bedrock Agents and Azure AI Foundry ship robust governance natively, the infrastructure-layer argument shifts decisively toward public cloud, and Broadcom's private cloud positioning becomes a niche play for regulated industries with data sovereignty requirements.
The Standardization Gambit
The AuthZEN integration deserves more attention than it has received. By adopting an open authorization standard, Broadcom is positioning itself to influence the emerging standards landscape for agent identity. If AuthZEN becomes the de facto protocol for agent authorization, Broadcom gains a seat at the table where the rules get written—a position that matters more than any individual product win.
This is the quiet play beneath the product announcement. The battle for AI agent governance is not just about features and performance. It is about who defines the protocols that all agents will use to authenticate, authorize, and audit their actions. Standards capture compound value over time, and the early movers in standards development accrue advantages that feature competition cannot overcome.
The Infrastructure Commitment Question
For enterprises evaluating AgentMinder, the fundamental question is not whether the product works—Broadcom's internal metrics suggest it does—but whether the infrastructure commitment is justified. AgentMinder is not a standalone product; it is a component of VMware Private AI Cloud, bundled with VCF 9, AI Factory, Tanzu services, vDefend, and Avi load balancers. Adopting it means deepening VMware commitment at a moment when many enterprises are diversifying across clouds.
The counterargument is that governance is precisely the kind of capability that benefits from deep infrastructure integration. Surface-level controls that can be bypassed by a determined agent are worse than no controls at all, because they create false confidence. The question enterprises must answer is whether their agent deployments are critical enough to warrant infrastructure-layer governance, or whether identity-layer controls provide sufficient coverage at a fraction of the implementation cost.
The Road Ahead
The AI agent governance market is forming in real time, and the next eighteen months will determine its structure. The signals to watch are concrete: whether Broadcom releases a standalone version for non-VMware environments, whether Okta and other identity providers extend their agent governance depth, whether cloud providers ship native governance that makes third-party solutions redundant, and whether AuthZEN gains adoption beyond its initial proponents.
The deeper question is whether governance becomes a feature of the infrastructure layer, the identity layer, or a new layer entirely. My read, based on the infrastructure patterns I have studied across payment rails and settlement systems, is that the layers will blur. Identity providers will add enforcement depth. Infrastructure vendors will add identity breadth. The winning architectures will be those that recognize governance as a cross-cutting concern requiring coordination across layers, not containment within one.
For now, Broadcom has made the most substantive infrastructure-layer argument in the market. Whether that argument wins depends less on the product itself than on the standards battle, the cloud provider response, and the willingness of enterprises to commit to the VMware stack at a moment of architectural transition. The control plane is built. The question is who gets to define the protocols that run through it.