Agent Gateway Tool-Level Access Policies for the Agentic Architect Exam

GCP Study Hub
November 18, 2026

Agent Gateway parses MCP traffic, which allows its egress access policies to govern individual tools on an MCP server rather than only allowing or blocking the entire server. On the Professional Agentic Architect exam, this capability matters wherever tool access has to be decided per agent and enforced by the platform rather than by the agent itself. Tool-level egress rules are bound to an agent's identity at the gateway, so the same server can offer different tools to different agents.

How tool-level egress policies work

Egress access policies on an Agent-to-Anywhere gateway are evaluated for the calling agent's identity, which agents receive through Agent Identity. Because the gateway understands MCP, a policy does not have to treat a server as a single destination. It can allow or deny individual tools, so an agent's access to a server can be narrower than the full set of tools that server exposes.

Policy conditions can test attributes of the tool being called. The iap.googleapis.com/mcp.tool.isReadOnly attribute matches tools marked as read-only on the MCP server, so a condition on it limits an identity to read-only operations. Each policy applies only to the identity it is written for, so narrowing one agent's tool access leaves the policies of every other agent unchanged.

Enforcing boundaries outside the agent's code

Enforcing tool access at Agent Gateway places the control outside the agent's code. Because the gateway evaluates the egress access policy for the agent's identity, nothing in the agent's code can widen the tools it is allowed to call. That separation satisfies security requirements that call for an enforced control independent of how the agent is implemented.

Per-identity gateway policies also keep permissions out of the server architecture. Splitting a server's tools across separate deployments according to who may call them is one way to vary access, and every split adds another server to run. With policies at the gateway, one deployment can serve agents with different permissions, and each identity reaches only the tools its policy allows.

Comparing Agent Gateway policies with other controls

Several Google Cloud and agent-level controls restrict behavior in different ways, and telling them apart comes up regularly when designing agent governance for the Professional Agentic Architect exam. Each alternative operates at a different layer from tool-level egress policies in Agent Gateway.

  • tool_filter in McpToolset: A tool filter in an agent's McpToolset narrows the available tools inside the agent's code. Because the filter lives in code, a code change could undo it, so it does not provide enforcement outside the agent's code.
  • System instructions: Instructing a model to call only read-only tools relies on the model following that text. That makes a system instruction guidance rather than an enforced control.
  • Project-level IAM roles: An IAM role on a Google Cloud project governs access to Google Cloud resources. It does not control access to individual tools on an internal MCP server.
  • VPC Service Controls: A VPC Service Controls perimeter works at the level of Google Cloud APIs. It cannot see the names of individual MCP tools, so it cannot allow some tools and block others on the same server.
  • Model Armor: Model Armor screens the content of prompts and tool calls. It evaluates content rather than deciding which tools an agent identity may call.

Choosing the right control layer

We would lean toward an Agent Gateway egress access policy whenever access has to be decided per agent identity and per MCP tool. Two questions narrow down the right control layer: whether enforcement must happen outside the agent's code, and whether the rule governs individual MCP tools rather than Google Cloud APIs or request content.

Model Armor fits when the goal is screening prompt and tool-call content. IAM roles and VPC Service Controls apply to Google Cloud resources and Google Cloud APIs. Rules about which tools an identity may call belong in Agent Gateway egress access policies, where a condition on iap.googleapis.com/mcp.tool.isReadOnly can limit an identity to tools marked as read-only.

Our Professional Agentic Architect practice exams test Agent Gateway tool-level access policies alongside Agent Identity and Model Armor, with detailed explanations that drill these distinctions.

Get tips and updates from GCP Study Hub

arrow