agent-platform-architecture

Designs, engineers, and validates enterprise Gemini platform architectures with multi-project hierarchies, governance, and security policies.

23.7k|2.0k|Updated Mar 31, 2026
Ready to run
npx skills add https://github.com/google/skills --skill agent-platform-architecture-google
💬 Prompt for ChatGPT / Claude Chat▼
Please help me install this Agent Skill.
Skill: agent-platform-architecture
Source: https://github.com/google/skills/tree/main/skills/cloud/agent-platform-architecture
Command: npx skills add https://github.com/google/skills --skill agent-platform-architecture-google

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Enterprises deploying Gemini platform agents face complex architectural decisions across project hierarchies, network isolation, identity governance, and policy enforcement. This Skill eliminates guesswork by providing systematic design patterns and production-ready implementation guidance. ## Core Features & Use Cases - Multi-Project Topology Design: Define Host, Governance Hub, Runtime Spoke, and Tool Spoke project hierarchies with VPC-SC and custom routing. - Agent Gateway and ACT Configuration: Choose between ALL_TRAFFIC and PRIVATE_RANGES_ONLY egress modes, configure Agent Connectivity Templates, and set up network attachments. - Security Policy Implementation: Deploy Unified Access Policy (UAP) V2, Semantic Governance Policies (SGP), and Model Armor content authorization. - Use Case: Your organization needs to deploy 50 agents across production environments with strict egress controls, private API routing, and role-based access. This Skill guides you through the complete architecture, from network design through Day-2 operations including ACT updates and agent lifecycle management. ## Quick Start Read the Domain Router and Symptom Matrix in SKILL.md to identify your architecture domain, then consult the corresponding reference manual (topologies_and_project_layouts.md, agent_gateway_and_act.md, egress_routing_and_swp.md, ingress_patterns_and_rewrites.md, governance_security_policies.md, registry_and_federation.md, or agent_runtime_and_identity.md) for exact CLI commands and YAML schemas.

Frequently Asked Questions about agent-platform-architecture

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I choose between ALL_TRAFFIC and PRIVATE_RANGES_ONLY agent connectivity modes?▼

Use ALL_TRAFFIC for regulated environments requiring 100% egress governance through Agent Gateway inspection, SWP filtering, and Cloud NAT static IPs. Use PRIVATE_RANGES_ONLY for dev/test where only RFC 1918 internal routing is monitored and external traffic bypasses the gateway.

What is a Policy-Based Route (PBR) and when do I need it for agent egress?▼

PBR routes traffic based on source IP ranges rather than network tags, solving the PSC Network Attachment tag incompatibility. Use priority-100 PBR to intercept PSC-I subnet traffic to the Secure Web Proxy internal IP, with priority-50 rules bypassing PBR for RFC 1918 and private Google API ranges.

How do I prevent an Agent Runtime instance deletion from failing with HTTP 400 FAILED_PRECONDITION?▼

Append ?force=true to the DELETE request to drop active sessions cleanly. Always delete ReasoningEngine instances before Agent Gateways, wait 20 seconds for control-plane settlement, then delete ACTs and wait 180 seconds for PSC-I NIC detachment before deleting Network Attachments.

What is the difference between Unified Access Policy V2 and legacy IAM Policy V1?▼

UAP V2 uses CEL expressions matching destination attributes (destination.is_registered, destination.agent_registry.mcp_server.tool.name) and applies hierarchically across Organization, Folder, and Project. Policy V1 uses legacy attribute conditions and evaluates only at the Resource Manager node level.

How do I configure Semantic Governance Policy (SGP) to inspect LLM payloads without blocking binary gRPC streams?▼

Set CONTENT_AUTHZ policyProfile with a when CEL clause excluding application/grpc content types: (!('content-type' in request.headers) || !request.headers['content-type'].lowerAscii().startsWith('application/grpc')). Register Spanner and OTLP endpoints with GRPC protocol binding, never HTTP_JSON.

Why does my Agent Runtime fail to initialize MCP tools with HTTP 403?▼

The UAP rule missing empty tool.name check. MCP discovery (initialize, tools/list) sets tool.name to empty string. Add (destination.agent_registry.mcp_server.tool.name == '' || destination.agent_registry.mcp_server.tool.name == 'SPECIFIC_TOOL') to all MCP allow rules.