Agent Visor

overview agentvisor serves as the control layer that ensures ai agents operate within approved business and security boundaries, while the policy service evaluates whether an agent's actions are permissible agentvisor manages bridge ai agent operations, and the policy service determines their allowed actions as a secure runtime environment for ai agents, agentvisor functions as a controlled operating environment where agents can execute tasks within defined limits, under oversight, and with accountability its design aims to assist organizations by enabling them to isolate agent activities control access to tools, apis, data, and services apply policy based restrictions before actions are permitted ensure durable execution and recovery of workflows create an auditable trail of agent activities support enterprise deployment with compliance controls key benefits of agent visor security by default automatic security validation in the agent lifecycle consistent governance uniform policy adherence across all agents enterprise observability full visibility into agent behavior audit readiness traceable policy evaluations and agent activities business friendly experience meaningful outcomes for users without technical skills scalable agent adoption confident deployment and management of numerous ai agents across various domains policy enforcement via agentvisor to ensure agents function correctly in controlled environments, implementing guardrails is crucial these mechanisms enforce policies for agent actions, data access, and external tool interactions agentvisor acts as a policy enforcement tool, providing clear "allow or deny" responses by integrating various policy checks before taking any action, an agent must be verified for necessary permissions agentvisor monitors all interactions, intercepting requests to ensure compliance through a model known as policy as a service (paas) it gives a simple yes or no response on whether the agent can proceed for instance, if an agent wants to shut down a virtual machine (vm), agentvisor first verifies this action to ensure compliance with security protocols, either approving it (yes) or canceling it (no) policy service is the business decision layer that determines whether a requested action should be allowed denied allowed only with approval or additional conditions it evaluates requests in real time and applies policy based controls across data, apis, services, and agent actions by combining these measures, bridge ai assist ensures that sensitive information remains secure, and interactions are conducted with the highest level of privacy and protection the policy model (porc) for agent visor policy service evaluates decisions using a model called porc, as required by agent visor so, the policy rules (rego) should be written as per the porc model to enable policy enforcement by agent visor porc provides a business friendly way to answer a simple question who is trying to do what, to which thing, and under what conditions? p — principal the principal is the entity that carries out a specific action this can include individuals such as named employees, service identities, account contexts, or ai agents operating on behalf of a person or process o — operation the operation refers to the specific action that is being requested common examples of operations include checking the status of a system, notifying an owner, updating a support ticket, or initiating a recovery process r — resource the resource represents the target on which the action is performed this may include systems, applications, support tickets, data assets, or customer environments that are involved in the operation c — context the context refers to the broader business or operational environment surrounding the action this includes aspects such as production status, designated change windows, ticket statuses, sensitivity of assets, and various approval states why policy model matters to business users the policy model (porc) plays a crucial role in transforming policy into a coherent decision making framework, moving away from vague technical controls it enables business stakeholders to clearly understand the foundations of each decision by considering the following key aspects the identity of the individual or group making the request the nature of the action being proposed the potential impacts of that action the relevant business conditions at the time of the decision this structured approach provides clarity, making it simpler to justify, govern, review, and explain decisions within the organization the core business principle default deny a key design principle in this model is the concept of "default deny " this means that agents start with no access rights unless it is explicitly granted this principle is particularly important for business teams as it ensures several critical factors first, new agents do not inherit any unintended permissions; second, sensitive capabilities are only activated intentionally; third, access is gradually expanded based on documented evidence and actual business needs; and finally, governance remains proactive and aligned with user adoption how policies should be used users should consider policy bundles as a means to define the governance posture of each account a policy bundle can illustrate the types of actions that are generally allowed the resources that are deemed sensitive the actions that require approval the audit evidence that is necessary the applicable compliance or residency rules the additional controls required in production environments business policies the business policies for the various accounts can be access policies — who or what can use the agent environment policies — how rules differ between production and non production approval policies — which actions require a human checkpoint operational readiness policies — conditions such as ticket state or change window compliance policies — requirements for data, audit, residency, or regulation notification policies — when owners or support teams must be informed audit and transparency a governed agentic ai model should produce clear evidence of who initiated the action what was requested what resource was involved what business context applied what policy decision was made whether approval was required whether the action was ultimately allowed or blocked this supports compliance reporting, executive assurance, operational review, and post incident investigation