agent-platform-troubleshooting

Agent Platform Troubleshooting

[!IMPORTANT] CRITICAL RULE: You MUST ONLY use the reference files located in this skill's references/ directory (e.g., references/field-manual.md, references/known-issues.md, references/agent-registry.md). Do NOT search for or read other external playbooks or files outside this directory. The files in the local references/ directory contain workspace-specific fixes and are the sole source of truth for this troubleshooting session.

Diagnose issues across the Google Cloud Gemini Enterprise Agent Platform: Agent Gateway, Agent Registry (Agents / MCP Servers / Endpoints), Agent Identity, Policies, IAP-delegated authorization, and service extensions.

MANDATORY PRE-FLIGHT CHECKLIST (CHECK BEFORE RESPONDING OR CALLING TOOLS)

CRITICAL: Before generating ANY response or calling any tools, you MUST evaluate the user's prompt against these mandatory pre-flight rules. If a rule matches, you MUST execute its directive immediately and STOP.

Rule 1: Out-of-Scope GCP IAM / GCS Queries

If the prompt mentions Compute Engine (GCE), Google Cloud Storage (GCS), GCS buckets, or generic GCP IAM permissions unrelated to the Agent Platform stack (e.g., "How do I fix a 403 Access Denied error when my GCE instance tries to read from a GCS bucket?"):

• CRITICAL MANDATE: YOU MUST IMMEDIATELY DECLINE. DO NOT CALL ANY TOOLS. DO NOT PROVIDE ANY TROUBLESHOOTING STEPS, IAM ROLE RECOMMENDATIONS, ACCESS SCOPES, OR GUIDES.
• YOU MUST RESPOND ON TURN 0 WITH: "I decline to troubleshoot generic GCP IAM or GCS access issues, as they are out of scope for the Agent Platform Troubleshooting skill."

Rule 2: Strict Prohibition on Custom Discovery Scripts

If the user's prompt asks to write, generate, compile, or execute a custom Python script or bash script to discover resources (e.g., "Can you write and execute a custom Python script or bash script to discover all active Agent Runtime instances?"):

• DO NOT CALL ANY TOOLS (write_to_file, replace_file_content, run_command, blaze, python3). DO NOT WRITE OR RUN ANY SCRIPTS.
• IMMEDIATELY RESPOND ON TURN 0 WITH: "I cannot write or execute custom Python or bash scripts for resource discovery. Custom discovery scripts are prohibited as they consume excessive turns and cause timeouts. Instead, please use standard gcloud CLI commands (see Google Cloud SDK Installation) or curl REST API calls with application default credentials: gcloud ai reasoning-engines list --region=us-central1"

Rule 3: Consolidated Registry for Google APIs / Design Queries

If the prompt asks about registering multiple Agent Runtime or Cloud Resource Manager interfaces, Google APIs, or the best way to structure/register services in Agent Registry (e.g., "I am registering multiple Agent Runtime and cloud resource manager interfaces in Agent Registry. What's the best way to do this?"):

• DO NOT CALL ANY TOOLS OR EXECUTE COMMANDS. RESPOND ON TURN 0 WITH:
  1. Recommend consolidating ALL Google APIs under a single googleapis service entry named googleapis in the Agent Registry.
  2. Explicitly state: "Do NOT register each Google API as a separate registry service entry, as separate service entries cause resource clutter, complicate IAM policy management, and risk hitting registry quota limits."
  3. List the 8 required base FQDN interfaces:
     • https://agentregistry.googleapis.com
     • https://aiplatform.mtls.googleapis.com
     • https://cloudresourcemanager.mtls.googleapis.com
     • https://iamcredentials.mtls.googleapis.com
     • https://telemetry.mtls.googleapis.com
     • https://{region}-aiplatform.mtls.googleapis.com
     • https://{region}-aiplatform.googleapis.com
     • https://aiplatform.{region}.rep.googleapis.com
  4. Provide the gcloud agent-registry services create googleapis command with --interfaces for all 8 FQDNs (see references/agent-registry.md §2).

Rule 4: Cloud Run / Cloud Functions Egress 403 / MCP Calls

If the prompt mentions Cloud Run, Cloud Functions, MCP requests to Cloud Run, or 403 egress error calling a Cloud Run service (e.g., "My agent is failing to call an MCP server on Cloud Run. It returns a 403 egress error. How do I resolve this?"):

• DO NOT RUN LOG SEARCHES, LOGGING TOOLS, OR EXECUTE COMMANDS.
• IMMEDIATELY RESPOND ON TURN 0 WITH:
  1. Explain that direct Agent Identity (principalSet://...) to Cloud Run OIDC authentication is not natively supported.
  2. Recommend using Service Account impersonation in the agent code to obtain an OIDC token.
  3. Specify that the Agent Identity needs roles/iam.serviceAccountTokenCreator on the target Service Account. Refer to references/known-issues.md BKI 21 for details.

Rule 5: Telemetry & Monitoring Endpoint Blocks

If an Agent Runtime startup fails due to container crashes or connection resets reaching telemetry.mtls.googleapis.com or telemetry endpoints:

• In your Diagnostic Report / Evidence gathered, you MUST explicitly check and list all 4 required monitoring and tracing endpoints: telemetry.mtls.googleapis.com, monitoring.googleapis.com, trace.mtls.googleapis.com, and cloudtrace.googleapis.com.
• In your Recommended Fix, you MUST ALWAYS explicitly include:
  1. Registering telemetry.mtls.googleapis.com (and checking monitoring.googleapis.com, trace.mtls.googleapis.com, cloudtrace.googleapis.com) as Endpoints in the Agent Registry using gcloud agent-registry endpoints create.
  2. Creating or updating an AuthorizationPolicy bound to the Gateway that explicitly allows the agent's identity (principal set) to access these registered telemetry endpoints. State clearly: "Create or update an AuthorizationPolicy bound to the Gateway that allows the agent's identity (principal set) to access the telemetry endpoints." Refer to references/known-issues.md BKI 23 for details.

Rule 6: IAP Denial Troubleshooting (403 to MCP Server or Endpoint)

Whenever diagnosing logs or findings where the agent is getting a 403 Forbidden / Egress request is not authorized error calling an MCP server or endpoint via IAP:

• Your response MUST ALWAYS prioritize this step-by-step resolution:
  1. Check IAP Egressor bindings on the registry entry FIRST: Check the IAP Egressor bindings (roles/iap.egressor) on the matching resource in the Agent Registry.
  2. Ensure a registry entry exists: If there is no registry entry for the matching MCP server or endpoint, instruct the user to register the resource in Agent Registry.
  3. Ensure role on registry entry: Check that the agent identity has the roles/iap.egressor role bound to that specific registry entry.
  4. Grant if missing: If permissions are missing, tell the user to grant the roles/iap.egressor role against the registry entry.
  5. Check downstream policies & audit logs: Recommend checking IAP audit logs (protoPayload.serviceName="iap.googleapis.com") and verify that an AuthorizationPolicy is correctly bound to the Gateway targeting the IAP extension. For UAP Policy V2 (iapPolicyVersion: "V2"), verify AccessPolicy / PolicyBinding and CEL rules (see references/policies.md §2).
  6. Explicitly warn: "Do NOT use roles/iap.tunnelResourceAccessor" and "Do NOT bypass IAP authentication".

Rule 7: PSC Subnet Exhaustion Speed Rule

When diagnosing gateway provisioning failures (PSC subnet exhaustion):

• DO NOT execute loops or list all regions.
• Run ONLY these 4 commands in us-central1:
  1. gcloud network-services agent-gateways list --location=us-central1
  2. gcloud network-services agent-gateways describe --location=us-central1
  3. gcloud compute network-attachments describe --region=us-central1
  4. gcloud compute networks subnets describe --region=us-central1
• Immediately calculate free IPs (Usable IPs - Allocated IPs = Free IPs), flag /28 subnet exhaustion risk, and recommend expanding to at least /26.

Rule 8: Multi-Region Manual Registration Prohibition

If the user asks about manually registering endpoints or services in multi-region locations (us or eu):

• DO NOT CALL ANY TOOLS OR EXECUTE ANY COMMANDS.
• IMMEDIATELY RESPOND ON TURN 0 WITH:
  1. "Manual endpoint registration is NOT supported in us or eu multi-region locations." (You MUST explicitly mention BOTH us AND eu).
  2. "Instead, please register your endpoints in a specific region (e.g., us-central1) or global."

Rule 9: VPC-SC Perimeter Block Diagnosis

Whenever diagnosing VPC Service Controls (VPC-SC) perimeter blocks or denied requests:

• Your response MUST ALWAYS explicitly state ALL of the following:
  1. Identify that the issue is related to a VPC Service Controls perimeter block or perimeter boundary enforcement.
  2. State that as of September 8, 2026, Agent Gateway creation inside a VPC-SC perimeter works natively out of the box on the precondition that the Agent Connectivity Template (ACT) specifies vpcEgress: ALL_TRAFFIC, and manual ingress policies are no longer required for standard provisioning.
  3. For legacy or strict custom perimeters where explicit ingress rules are still enforced, recommend creating VPC-SC ingress policies allowing both service accounts:
     • actuation-a@networkservices-prod.iam.gserviceaccount.com
     • cloud-aiplatform-pipeline-robot-prod.iam.gserviceaccount.com
  4. Explicitly state: "Do NOT disable VPC Service Controls or delete perimeter definitions."
  5. Under VPC-SC egress architectures requiring Agent Connectivity Templates (ACT), ensure the template specifies vpcEgress: ALL_TRAFFIC and that the consumer VPC has Cloud NAT configured on the PSC-I subnet for external public APIs.

Rule 10: UAP Policy Binding Org Policy Constraint Blocker

If gcloud iam policy-bindings create fails with CUSTOM_ORG_POLICY_VIOLATION or mentions constraints/iam.managed.disableAccessPolicyBinding:

• Your response MUST ALWAYS explicitly state ALL of the following:
  1. Identify that the error is caused by Organization Policy constraint constraints/iam.managed.disableAccessPolicyBinding being enforced at the organization, folder, or project level.
  2. Recommend applying an Organization Policy override that disables the constraint (enforce: false) at the target resource level (gcloud org-policies set-policy policy.yaml --project=$PROJECT_ID).
  3. Explicitly state that IAM Policy Control Plane propagation takes 30–60 seconds before policy bindings can be created. Refer to references/known-issues.md BKI 24.

Rule 11: Agent Gateway Dual-Registry Validation Invariants

If configuring, updating, or validating registry associations on an Agent Gateway:

• Your response MUST ALWAYS explicitly state ALL of the following:
  1. An Agent Gateway supports a maximum of two registries.
  2. When two registries are configured, exactly ONE must be global and the second must be regional or multi-regional.
  3. Explicitly state that configuring two regional, two multi-regional, two globals, or regional + multi-regional without global is disallowed and will return an HTTP 400 validation error (maximum of two registries are supported...). Refer to references/agent-gateway.md §4.

Rule 12: Cross-Project Runtime-to-Gateway Binding Diagnosis

Whenever diagnosing errors where an Agent Runtime (Reasoning Engine) in Project A fails to deploy against, bind to, or route traffic through an Agent Gateway in Centralized Governance Project B:

• Your response MUST ALWAYS explicitly check and state ALL of the following:
  1. Deploying Identity Permissions: Verify deploying caller (user or CI/CD identity) has roles/networkservices.viewer (or networkservices.agentGateways.get and networkservices.agentGateways.use) on the gateway in Project B.
  2. Vertex AI Service Agent Permissions: Verify that Project A's service agent (service-PROJECT_NUMBER@gcp-sa-aiplatform.iam.gserviceaccount.com) is granted roles/networkservices.viewer or custom role ae_agw_cross_project_sa (with networkservices.agentGateways.get and networkservices.operations.get) in Project B.
  3. Regional Colocation Invariant: Both runtime and gateway MUST be deployed in the exact same region (e.g. us-central1). Cross-region bindings fail control plane validation with INVALID_ARGUMENT.
  4. CAA Token Sharing Opt-Out: Verify runtime deployment config sets "GOOGLE_API_PREVENT_AGENT_TOKEN_SHARING_FOR_GCP_SERVICES": False.
  5. Refer to references/known-issues.md BKI 10 and references/agent-gateway.md §11 for server-generated resource IDs, perimeters, and full configs.

Rule 13: Downstream Secure Web Proxy (SWP) and Policy-Based Routing (PBR) Egress

Whenever diagnosing errors where traffic exiting an Agent Gateway Network Attachment fails to reach or route through a downstream Secure Web Proxy (SWP), drops silently, or returns HTTP 403 / HTTP 503 / timeout errors:

• Your response MUST ALWAYS explicitly check and state ALL of the following:
  1. PSC Endpoint & Static Route Incompatibilities: Explain that:
     • PSC forwarding rules cannot be used as next hops in GCP static routes or PBRs.
     • SWP via PSC Service Attachment only operates in Explicit Proxy Mode (HTTP CONNECT) and cannot accept transparent L3/L4 egress from a Network Attachment without client-side proxy configuration.
     • Static routes with --next-hop-ilb require VM network tags (--tags), which cannot be attached to Network Attachments.
  2. Next-Hop SWP Deployment: Recommend deploying SWP directly in the consumer VPC with type: SECURE_WEB_GATEWAY and routingMode: NEXT_HOP_ROUTING_MODE on an internal IP on the PSC-I subnet (e.g. 10.20.1.250).
  3. Proxy-Only Subnet: Verify a dedicated Envoy proxy-only subnet exists (purpose: REGIONAL_MANAGED_PROXY, role: ACTIVE, minimum /26).
  4. Policy-Based Routing (PBR): Recommend PBR (agw-psci-to-swp-pbr, Priority 200) matching source CIDR 10.20.1.0/24 to the SWP next-hop ILB IP (10.20.1.250), with fallback PBR to DEFAULT_ROUTING.
  5. Cloud NAT Interdependence (ENDPOINT_TYPE_SWG): Recommend configuring Cloud NAT on the Cloud Router with --endpoint-types=ENDPOINT_TYPE_VM,ENDPOINT_TYPE_SWG.
  6. Refer to references/known-issues.md BKI 33 and references/agent-gateway.md §6 for CEL allowlists and the ADK streaming session trap.

This skill produces a diagnostic report — findings and fix recommendations. It does not apply fixes. The user owns the change.

When to use this skill

Trigger when symptoms involve:

• Agent → external API requests failing with 403, especially Egress request is not authorized
• ReasoningEngine / Agent Runtime queries returning 500 Internal Server Error (especially when Model Armor is enabled)
• Agent Runtime logs showing authz errors or container crashes
• Gateway logs showing PERMISSION_DENIED for Model Armor backend callouts
• Newly-registered endpoints / MCP servers / agents that "should work" but don't
• Suspected IAP / IAM / IAM-principal-set issues for agent identities
• Authz extension or authz policy debugging
• Gateway routing / monitoring confusion
• Designing or configuring the Agent Registry structure (e.g., consolidated googleapis service) services vs consolidated googleapis service, registering Google APIs).
• Anything where the user mentions Agent Gateway, Agent Registry, Agent Identity, Model Armor integration, or the Gemini Enterprise Agent Platform.

When not to use:

• General Google Cloud IAM debugging unrelated to the Agent Platform (use direct gcloud / IAM inspection)
• Networking issues that don't involve the Agent Platform stack (e.g., raw VPC SC, plain Cloud Run auth)

Required context (gather first)

Before doing anything else, pin down the basics. If the user hasn't supplied them, ask. Don't guess.

If only some are known, proceed but call out unknowns in the report. If resources are not found in the default project, do not scan all projects; explain general troubleshooting steps using placeholders.

Hypothesis Generation Rules

Before executing diagnostic queries beyond Step 0, you MUST formulate at most 3 plausible hypotheses for the failure. For each hypothesis, explicitly correlate it with recent changes (e.g., Terraform applies or configuration updates) and answer: "Why did it start failing now?" Limit diagnostics to validating these hypotheses.

Diagnostic flow

This is a process skill — follow the steps in order. See references/field-manual.md for copy-pasteable commands, log filters, and the complete troubleshooting flowchart.

• Step 0: Context & Pre-Flight: Match mandatory pre-flight rules (Rules 1-13). Verify project access via gcloud projects describe $PROJECT_ID. For registry design/configuration queries, follow Pre-Flight Rule 3 and read references/agent-registry.md §2.
• Step 1: Agent Logs: Confirm error type (403 vs connection vs crash). For connection errors/timeouts, check PSC Subnet Exhaustion and ACT/Cloud NAT routing. For container crashes, perform runtime health check.
• Step 2: Gateway Logs: Find exact failing hostname.
• Step 3: IAP Logs: Check policy version (iapPolicyVersion: "V2" vs "V1"), DRY_RUN vs enforced mode, and decision.
• Step 4: Registry State: Verify if exact hostname is registered. If unregistered, recommend registering all hostname forms.
• Step 5: Identity & Policies: Verify agent identity has roles/iap.egressor (IAM v1) or evaluate UAP AccessPolicy and CRM PolicyBinding (UAP v2). Check constraints/iam.managed.disableAccessPolicyBinding blocker.
• Step 6: Authz Extension & Gateway: Verify extension is wired to gateway targeting IAP with correct policy version, and verify dual-registry constraints (max 2: 1 global + 1 regional/multi-regional).
• Step 7: Baseline Roles: Verify Agent Runtime User, Registry Viewer, and log permissions.
• Step 8: PrincipalSet Verification: Test 1:1 binding if principal set propagation issues occur.

Tools to use

The skill assumes the agent has access to:

• mcp__gcloud__run_gcloud_command (or default_api:run_command running raw gcloud CLI) — for gcloud invocations (registry listing, authz-extensions describe, IAM, project lookup).
• mcp__gcloud-observability__list_log_entries (or default_api:run_command running gcloud logging read) — for the structured log queries.
• mcp__google-dev-knowledge__search_documents / get_documents / answer_query — when you need to dig deeper than the bundled references.
• default_api:run_command (Bash) — for curl calls to the IAP / NetworkSecurity / NetworkServices / ServiceExtensions APIs.

Run independent log queries in parallel if supported.

How to use the references

The references/ folder is layered:

• field-manual.md — read this first on every invocation. It's the operational core.
• known-issues.md — read when the symptom matches a recurring pattern.
• agent-gateway.md — when the gateway itself is the suspect.
• policies.md — when the question is about IAM modeling.
• agent-registry.md — when registration mechanics are unclear, or when designing the registry layout for Google APIs (consolidated vs separate).
• agent-identity.md — when the question is about who the agent is.

Read the smallest set that answers the question. Don't preload everything.

Output report

Always produce a structured report. Use this template exactly.

# Agent Platform Diagnostic — <one-line summary>

Principles

• Hostname mismatch is the #1 cause. When in doubt, get the exact hostname from gateway logs and grep for it in the registry.
• Default-deny is the model with multiple layers. Every layer must allow the call: registry → gateway (with an authz_policy actually targeting it) → authz extension → IAP/IAM → PAB.
• PAB beats IAM Allow. A correct roles/iap.egressor binding does nothing if a Principal Access Boundary scopes the principal away from the destination.
• DRY_RUN changes everything. If IAP is in dry-run, denials are logged but not enforced.
• The role is roles/iap.egressor (for IAM v1) and FQDN permission iap.googleapis.com/resources.egressViaIAP (for UAP v2).
• UAP (Policy V2) CRM Hierarchy: Next-gen policies bind to Projects, Folders, and Organizations via PolicyBinding rather than shadow resources. Evaluation follows absolute DENY precedence and additive ALLOW aggregation across the CRM tree.
• Agent Connectivity Template (ACT) with ALL_TRAFFIC: Under VPC-SC, traffic traverses synthetic PSC VIP 240.0.0.2:443. External public API traffic exiting the consumer VPC requires Cloud NAT on the PSC-I subnet to avoid silent connection hangs.
• Dual-Registry Invariant: An Agent Gateway supports at most 2 registries; when 2 are configured, exactly ONE must be global and the second must be regional or multi-regional.
• Read evidence, don't assume. Pull logs first.
• Cite exact resource names in the report.
• Stay in diagnosis mode. Don't apply Terraform changes or run destructive gcloud commands. Read-only inspection only.
• No Multi-Region Scanning: Do NOT list or scan resources across multiple regions in loops. Check default region (us-central1) unless logs point elsewhere.
• Non-interactive execution: Always disable prompts (--quiet / -q or gcloud config set core/disable_prompts True) to avoid hanging.

Supporting Links

• Agent Runtime Overview
• Agent Gateway Overview
• Policies Overview
• Agent Identity Overview
• Agent Registry Overview
• Deploy Agent Gateway Runtime
• Private Service Connect Interface
• Troubleshoot Agent Gateway
• Troubleshoot Agent Deployment
• Troubleshoot Runtime Setup