Runtime Governance Architecture

Policy Decision Point vs Policy Enforcement Point for AI Agents

A Policy Decision Point (PDP) evaluates policy against contextual attributes and renders an authorization decision. A Policy Enforcement Point (PEP) intercepts the actual request and carries out that decision. For AI agents, this separation must extend to every tool call an agent makes, not just the initial user request, because agents generate multiple chained actions per task rather than a single discrete request.

Definitions Rooted in Established Access Control Models

The PDP/PEP model did not originate with AI agents. NIST's guidance on Attribute-Based Access Control describes the Policy Decision Point as the component that evaluates applicable policy and renders an authorization decision, and the Policy Enforcement Point as the component that intercepts a resource access request and carries out that decision. The OASIS XACML standard formalizes this into a reference architecture: a PEP forwards an access request to a PDP, the PDP evaluates policy against attributes of the subject, resource, and action, and the PEP enforces whatever permit or deny decision comes back.

This separation exists for a reason. Decision logic changes frequently as policy evolves, while enforcement infrastructure tends to be more stable and closer to the resource being protected. Keeping them distinct allows policy to be reviewed, tested, and updated without touching the enforcement layer, and it allows enforcement points to be added or moved without rewriting policy logic.

Two Distinct Functions in Agent Authorization

Policy Decision Point

Evaluates identity, requested action, tool parameters, and task state to render a permit or deny decision.

Policy Enforcement Point

Intercepts the agent's tool call or action and enforces the decision returned by the PDP.

Why AI Agent Runtimes Break the Traditional Assumptions

Traditional PDP/PEP architectures assume a single, discrete access request: a user asks for a resource, the PEP intercepts it, the PDP decides, the PEP enforces. AI agents do not operate this way. A single task can generate a chain of tool calls, each with different parameters, targets, and intent, executed in sequence or in parallel as the agent reasons through a problem.

This means enforcement cannot happen once per session. It has to happen at each tool or function-call boundary in the agent's execution path. It also means the PDP cannot rely on static role assignments alone. Because agent reasoning is non-deterministic, the same nominal action can carry different parameters across executions, so the PDP has to evaluate live runtime context, including the specific tool being invoked, the parameters passed to it, and the current task or session state, rather than a fixed permission set assigned in advance.

Agent identity adds another dimension. Agents frequently act on behalf of a user or another agent, which introduces delegated authority as an attribute the PDP must evaluate in addition to the traditional subject, resource, and action tuple.

PDP and PEP Responsibilities Compared

AspectPolicy Decision Point (PDP)Policy Enforcement Point (PEP)
Primary roleEvaluates policy against context and renders a permit or deny decisionIntercepts the request or tool call and carries out the decision
Trigger point in agent runtimeInvoked at every tool or function-call boundaryPositioned directly at each tool call interception point
Inputs evaluatedIdentity, delegated authority, tool parameters, task or session stateThe decision returned by the PDP
StabilityChanges frequently as policy evolvesTends to remain stable and close to the resource
Failure behaviorAmbiguous or delayed evaluation should not imply approvalShould default to denying the action if no decision is received in time

Where the Model Context Protocol Fits, and Where It Doesn't

The Model Context Protocol defines a client-server architecture in which AI models invoke external tools and resources through standardized requests. This creates discrete, interceptable call boundaries, which is useful for enforcement, since each tool call is a natural point at which a PEP can sit.

However, MCP does not itself define a standardized authorization or policy enforcement layer for tool calls. Authorization is left entirely to implementers. This is a common point of confusion: teams sometimes assume that adopting MCP implies some baseline of access control, when in fact the protocol only handles transport and message structure. The PDP and PEP have to be built or integrated separately, and enterprises should not treat protocol-level compliance as equivalent to policy enforcement.

Auditability Depends on Separating the Two Functions

When PDP decisions and PEP enforcement actions are logged together or not distinguished at all, it becomes difficult to determine whether an incident resulted from a policy gap or an enforcement failure. Logging the PDP's decision, including the context it evaluated and the reasoning attributes involved, separately from the PEP's enforcement action, including what was blocked or allowed and when, gives security teams the ability to reconstruct what happened after the fact.

This separation also supports least-privilege enforcement in a more granular way than role-based grants typically allow. A PDP evaluating narrowly scoped, context-specific permissions per tool call, rather than a broad role assigned once, more closely matches how OWASP's guidance on LLM applications frames the risk of excessive agency: agents invoking external functions should be checked for least-privilege scope and authorization at the specific point where the action is triggered, not upstream at a general session level.

Key implementation point

If the PDP does not respond in time, the PEP should default to denying the action rather than allowing it to proceed. An unavailable or ambiguous decision should never be treated as an implicit approval.

Evaluating Governance Approaches

Security engineers assessing how to implement PDP/PEP separation for agent runtimes should look at where enforcement actually sits relative to tool calls, what runtime attributes the PDP evaluates before deciding, and how decisions and enforcement actions are logged. Trussed AI provides runtime governance for enterprise AI agents, including runtime policy enforcement, agent identity and permissions, tool approval workflows, and audit logging designed to support this kind of separation and traceability across MCP and other agent tool interactions. These are architectural capabilities, not a substitute for defining the policy logic itself, which remains an organizational responsibility.

Frequently Asked Questions

Can a PDP and PEP be the same component in an AI agent system?

They can be co-located in a single service, but the functions remain logically distinct. Combining them into one undifferentiated process makes it harder to audit whether a failure was a policy evaluation error or an enforcement gap.

Does Model Context Protocol include built-in authorization?

No. MCP defines the structure and transport for tool calls but leaves authorization and policy enforcement to be implemented separately by the organization deploying the agent.

How often should the PDP be invoked during an agent task?

At every tool or function call, since agent tasks typically involve multiple chained actions rather than a single request, and permissions may need to be evaluated differently for each one.

What happens if the PDP does not respond in time?

The PEP should default to denying the action rather than allowing it to proceed, since an unavailable or ambiguous decision should not be treated as an implicit approval.

Assess Your Agent Authorization Architecture

Review where policy decisions are made and where they are enforced across your AI agent runtime before gaps in tool call authorization become audit or security incidents.

Explore Runtime Governance