Broadcom Launches AgentMinder to Govern Every Agent Tool Call with AuthZEN and OpenTelemetry

Broadcom announced AgentMinder at VMware Explore and made it generally available, seeking to shift agent governance from filtering model outputs to the actual execution path. The cloud-native gateway is deployed alongside models and enterprise backends, establishes auditable identities for agents, and determines—on a call-by-call basis—which backend should receive each tool call based on the declared task, permitted intent, tool, resource, user identity, and real-time risk. This is more granular than issuing an agent a single long-lived token at login, because the same agent can receive different decisions across tasks or contexts.

The product integrates with existing authorization systems through the OpenID Foundation’s AuthZEN Authorization API. The standard separates the policy enforcement point from the policy decision point and describes requests in terms of subject, action, resource, and context, allowing enterprises to retain their existing policy engines without routing all traffic through a single SaaS service. AgentMinder also uses OpenTelemetry to record sessions, actions, and chains of responsibility. It supports on-premises environments, VPCs, public clouds, and standard Kubernetes. Broadcom says its multi-region active-active deployment processes nearly 36 million customer-related and 7 million employee-related API calls per day, but these are internal production figures rather than reproducible throughput or latency benchmarks.

The biggest implementation question is whether governance coverage amounts to genuine isolation. If an agent can bypass the gateway through a direct connection, execute shell commands, or use unregistered credentials, per-call policies cannot form a complete security boundary. The public announcement also does not specify whether the system fails open or fails closed when the decision service is unavailable, how policy-update consistency is handled, whether OpenTelemetry fields contain sensitive prompts, or the extent of parameter-level mapping for MCP and A2A. Engineering teams should next request a public API specification and threat model, measure the tail latency added by each authorization decision, and validate the security claims with tests covering bypasses, replay attacks, cross-agent delegation, and policy-service outages.