
A principal access boundary (PAB) policy defines which resources a set of principals is eligible to access. Even when an IAM allow policy grants a broader role, the boundary policy still applies, so a mistaken grant on an outside resource has no effect. That behavior makes a PAB policy a guardrail against future IAM mistakes rather than a fix for current ones. On the Professional Agentic Architect exam, PAB policies appear in access control scenarios where teams need to limit which projects deployed agents can reach, regardless of how roles might be granted later.
When an organization runs multiple agents with Agent Identity inside a project, administrators often need to keep all of those agents within a specific set of approved projects. To establish that boundary, you create a PAB policy listing the approved projects and bind it to the principal set representing every agent identity in the project where the agents live.
The principal set for all agent identities in a project takes this form:
//agents.global.org-ORG_ID.system.id.goog/attribute.container/projects/PROJECT_NUMBER
Binding the PAB policy to this principal set restricts every agent identity in that project to resources in the listed approved projects. If an allow policy later grants one of those agents a role on a resource in an unrelated project by mistake, the boundary still applies and the grant has no effect. Because agents with Agent Identity do not run as service accounts, binding the boundary to this agent identity principal set ensures the policy actually covers the agents, whereas policies written for service accounts would miss them.
Several other IAM mechanisms, organization controls, and operational reviews overlap with access governance on Google Cloud, but each addresses a different problem than limiting which resources a set of agent identities is eligible to reach.
When evaluating access control options for the Professional Agentic Architect exam, it helps to separate what each control restricts. Deployment constraints govern where Agent Runtime can run, domain restricted sharing governs which identities may receive roles, and periodic reviews detect mistakes only after they have happened. Deny policies either fail to scale across unrelated projects, miss Agent Identity principals when aimed at service accounts, or block required projects when applied across the entire organization. When the requirement is a guardrail that keeps every agent identity in a project limited to approved projects even if broader allow policies are granted later, we would lean toward binding a principal access boundary policy to the project's agent identity principal set.
Our Professional Agentic Architect practice exams test principal access boundary policies alongside Agent Identity and IAM deny policies, with detailed explanations that drill these distinctions.