AI Agent Identity Governance: A Least-Privilege Checklist Before You Deploy

Broadcom's AgentMinder, part of the VMware AI Factory announced August 31, 2026, treats AI agents as enterprise-grade identities — binding each one's authority to a specific mission, a defined set of approved tools, and authorized resources, enforced at runtime. It's a genuinely useful architectural pattern, and its arrival is itself a signal: most organizations running AI agents today have no equivalent discipline in place, regardless of which vendor's tooling they use. The gap AgentMinder is built to close already exists in most environments right now.

Why Agent Permissions Tend to Sprawl Silently

An AI agent typically inherits the credentials of whatever service account or API key it was set up under during a proof-of-concept, and that scope rarely gets revisited once the agent moves toward production. Unlike a human employee's access, which usually gets reviewed at least occasionally through onboarding, role changes, or offboarding processes, an agent's permissions can sit unexamined indefinitely — a gap made worse by how easy it is to spin up a new agent quickly without going through the same access-request process a new employee would.

A Least-Privilege Checklist You Can Apply Today

• Write down the specific mission each agent is authorized to perform, in plain language — if you can't state it in one sentence, its scope is probably too broad to reason about
• Enumerate the exact tools and resources each agent can access, and compare that list against what its actual mission requires — anything extra is unscoped risk sitting idle until something goes wrong
• Check whether agent credentials are shared across multiple agents or use cases — shared credentials make it impossible to audit which agent did what, and mean one compromised agent's blast radius extends to everything sharing its identity
• Confirm agent actions are logged with enough detail to reconstruct what happened after the fact — 'the agent ran' is not sufficient audit trail; you need which tool, which resource, and what mission it was operating under
• Set a recurring review cadence for agent permissions, the same way you would for human access — an agent's scope should shrink over time toward exactly what it needs, not silently grow as new integrations get added

You Don't Need AgentMinder Specifically to Start

The principle behind AgentMinder — treat agents as first-class, mission-scoped identities with auditable, least-privileged access — is implementable with existing identity and access management tooling most enterprises already run, even before adopting any specific new platform. The checklist above is the actual work; a dedicated governance product makes it easier to enforce at scale, but the underlying discipline is available to start today, with whatever agents you already have running.