Broadcom launches AgentMinder to govern autonomous AI agents in enterprises

On 31 August 2026, at VMware Explore in Las Vegas, Broadcom announced AgentMinder, a management platform that gives enterprises control over autonomous AI agents. Where earlier AI tools primarily generate text, modern agents independently access data, call external services and execute business processes, requirements that place different demands on security and control.

AgentMinder is intended as an intermediate layer that brings together identity, access policy and observability. Broadcom positions it as a complement to existing security infrastructure, not a replacement for it. The platform was generally available from the date of announcement.

Why existing security models fell short

Traditional access control is built around human users: a person logs in, receives rights based on their role and acts within predictable patterns. An autonomous AI agent operates differently. Within a single session it can make dozens of tool calls, switch context and make decisions without direct human intervention, a poor fit for conventional security layers.

Risks that arise include unintended access to restricted data, unauthorised changes to systems, or an agent exceeding the boundaries of authority within which it is expected to operate. AgentMinder is designed to make those risks manageable without requiring organisations to rebuild their entire authorisation stack.

Three functional layers

AgentMinder operates through three layers that together cover the lifecycle of an agent session.

• Identity and intent. Each agent is assigned an enterprise identity and must declare at the start of a session what its purpose is, which tools it is permitted to use and which resources it has access to. Access is therefore tied not only to who the agent is, but also to what it intends to do.
• Runtime enforcement. A cloud-native AI gateway evaluates every individual tool call against established policy and current context, including the identity of the end user who initiated the agent. The system applies the least-privilege principle: an agent receives only the rights it needs at that moment.
• Observability and audit. AgentMinder is built on OpenTelemetry and captures full telemetry per agent session, including audit trails, anomaly detection and what Broadcom describes as chain-of-custody tracking. The latter makes it possible to trace the steps an agent took before a particular outcome was produced.

Integration with existing authorisation systems is handled via the AuthZEN standard, allowing organisations to reuse existing policy enforcement points.

How Broadcom uses it internally

Broadcom did not look to an external reference to demonstrate the platform's capabilities: the company already runs AgentMinder internally at scale. The system processes nearly 36 million customer-related and approximately 7 million employee-related API calls each day. More than 20 million customer identities and 72,000 employee identities are managed by the platform.

The architecture is set up as active-active across multiple regions, enabling continuous availability. Broadcom CIO Alan Davidson states in materials surrounding the launch that the system combines global scale with zero downtime during maintenance. That internal track record forms part of how Broadcom presents the platform to enterprise customers.

Deployment and infrastructure requirements

AgentMinder runs in Kubernetes-based environments. Broadcom cites VMware vSphere Kubernetes Service and Google Cloud Platform as supported platforms. Deployment is possible on-premises, in a virtual private cloud or in a public cloud. The choice of a cloud-native architecture aligns with how most large enterprises have structured their infrastructure in recent years.

Clayton Donley, Vice President and General Manager of Broadcom's Identity Management Security Division, states in a press release that AgentMinder enables companies to run their operations with confidence, even as the number of active AI agents grows.

What this means for those building with AI or investing in it

AgentMinder is an example of a category of products that is rapidly gaining importance: platforms that do not deliver the AI model itself, but the infrastructure built on top of it. As companies deploy more agents for processes in finance, HR and IT, demand grows for oversight mechanisms that demonstrably work and are auditable, particularly in sectors with strict compliance requirements.

For builders of AI agents and orchestration platforms, the choice of standards such as OpenTelemetry and AuthZEN is relevant: these partly determine how well a governance layer integrates with an existing stack. The fact that Broadcom, with its enterprise infrastructure background through VMware, is entering this segment signals that agent governance is being taken seriously as an infrastructure problem, beyond the circles of AI-native startups as well.