Vertex Engine Inspector
Overview
Inspect and validate Vertex AI Agent Engine deployments across seven categories: runtime configuration, Code Execution Sandbox, Memory Bank, A2A protocol compliance, security posture, performance metrics, and monitoring observability. This skill generates weighted production-readiness scores (0-100%) with actionable recommendations for each deployment.
Prerequisites
gcloud CLI authenticated with roles/aiplatform.viewer and roles/monitoring.viewer IAM roles- Access to the target Google Cloud project hosting the Agent Engine deployment
curl or gcloud for A2A protocol endpoint testing (AgentCard, Task API, Status API)- Cloud Monitoring API enabled for performance metrics retrieval
- Familiarity with Vertex AI Agent Engine concepts: Code Execution Sandbox, Memory Bank, Model Armor
Instructions
- Connect to the Agent Engine deployment by retrieving agent metadata via
gcloud ai agents describe
- Parse the runtime configuration: model selection (Gemini 2.5 Pro/Flash), tools enabled, VPC settings, and scaling policies
- Validate Code Execution Sandbox settings: confirm state TTL is 7-14 days, sandbox type is
SECURE_ISOLATED, and IAM permissions are scoped to required GCP services only
- Check Memory Bank configuration: verify enabled status, retention policy (min 100 memories), Firestore encryption, indexing enabled, and auto-cleanup active
- Test A2A protocol compliance by probing
/.well-known/agent-card, POST /v1/tasks:send, and GET /v1/tasks/<task-id> endpoints for correct responses
- Audit security posture: validate IAM least-privilege roles, VPC Service Controls perimeter, Model Armor activation, encryption at rest and in transit, and absence of hardcoded credentials
- Query Cloud Monitoring for performance metrics: request count, error rate (target < 5%), latency percentiles (p50/p95/p99), token usage, and cost estimates over the last 24 hours
- Assess monitoring and observability: confirm Cloud Monitoring dashboards, alerting policies, structured logging, OpenTelemetry tracing, and Cloud Error Reporting are configured
- Calculate weighted scores across all categories and determine overall production readiness status
- Generate a prioritized list of recommendations with estimated score improvement per remediation
See ${CLAUDE_SKILL_DIR}/references/inspection-workflow.md for the phased inspection process and ${CLAUDE_SKILL_DIR}/references/inspection-categories.md for detailed check criteria.
Output
- Inspection report in YAML format with per-category scores and overall readiness percentage
- Runtime configuration summary: model, tools, VPC, scaling settings
- A2A protocol compliance matrix: pass/fail for AgentCard, Task API, Status API
- Security posture score with breakdown: IAM, VPC-SC, Model Armor, encryption, secrets
- Performance metrics dashboard: error rate, latency percentiles, token usage, daily cost estimate
- Prioritized recommendations with estimated score improvement per item
See ${CLAUDE_SKILL_DIR}/references/example-inspection-report.md for a complete sample report.
Error Handling
| Error |
Cause |
Solution |
| Agent metadata not accessible |
Insufficient IAM permissions or incorrect agent ID |
Verify roles/aiplatform.viewer granted; confirm agent ID with gcloud ai agents list |
| A2A AgentCard endpoint 404 |
Agent not configured for A2A protocol or endpoint path incorrect |
Check agent configuration for A2A enablement; verify /.well-known/agent-card path |
| Cloud Monitoring metrics empty |
Monitoring API not enabled or no recent traffic |
Run gcloud services enable monitoring.googleapis.com; generate test traffic first |
| VPC-SC perimeter blocking access |
Inspector running outside VPC Service Controls perimeter |
Add inspector service account to access level; use VPC-SC bridge or access policy |
| Code Execution TTL out of range |
State TTL set below 1 day or above 14 days |
Adjust TTL to 7-14 days for production; values above 14 days are rejected by Agent Engine |
See ${CLAUDE_SKILL_DIR}/references/errors.md for additional error scenarios.
Examples
Scenario 1: Pre-Production Readiness Check -- Inspect a newly deployed ADK agent before production launch. Run all 28 checklist items across security, performance, monitoring, compliance, and reliability. Target: overall score above 85% before approving production traffic.
Scenario 2: Security Audit After IAM Change -- Re-inspect security posture after modifying service account roles. Validate that least-privilege is maintained (target: IAM score 95%+), VPC-SC perimeter is intact, and Model Armor remains active.
Scenario 3: Performance Degradation Investigation -- Inspect an agent showing elevated error rates. Query 24-hour performance metrics, identify latency spikes at p95/p99, check auto-scaling behavior, and correlate with token usage patterns to isolate the root cause.
Resources
1---2name: vertex-engine-inspector3description: Execute inspect and validate Vertex AI Agent Engine deployments including Code Execution Sandbox, Memory Bank, A2A protocol compliance, and security posture. Generates production readiness scores. Use when asked to "inspect agent engine" or "validate depl... Trigger with relevant phrases based on skill purpose.4license: MIT5---6# Vertex Engine Inspector78## Overview910Inspect and validate Vertex AI Agent Engine deployments across seven categories: runtime configuration, Code Execution Sandbox, Memory Bank, A2A protocol compliance, security posture, performance metrics, and monitoring observability. This skill generates weighted production-readiness scores (0-100%) with actionable recommendations for each deployment.1112## Prerequisites1314- `gcloud` CLI authenticated with `roles/aiplatform.viewer` and `roles/monitoring.viewer` IAM roles15- Access to the target Google Cloud project hosting the Agent Engine deployment16- `curl` or `gcloud` for A2A protocol endpoint testing (AgentCard, Task API, Status API)17- Cloud Monitoring API enabled for performance metrics retrieval18- Familiarity with Vertex AI Agent Engine concepts: Code Execution Sandbox, Memory Bank, Model Armor1920## Instructions21221. Connect to the Agent Engine deployment by retrieving agent metadata via `gcloud ai agents describe`232. Parse the runtime configuration: model selection (Gemini 2.5 Pro/Flash), tools enabled, VPC settings, and scaling policies243. Validate Code Execution Sandbox settings: confirm state TTL is 7-14 days, sandbox type is `SECURE_ISOLATED`, and IAM permissions are scoped to required GCP services only254. Check Memory Bank configuration: verify enabled status, retention policy (min 100 memories), Firestore encryption, indexing enabled, and auto-cleanup active265. Test A2A protocol compliance by probing `/.well-known/agent-card`, `POST /v1/tasks:send`, and `GET /v1/tasks/<task-id>` endpoints for correct responses276. Audit security posture: validate IAM least-privilege roles, VPC Service Controls perimeter, Model Armor activation, encryption at rest and in transit, and absence of hardcoded credentials287. Query Cloud Monitoring for performance metrics: request count, error rate (target < 5%), latency percentiles (p50/p95/p99), token usage, and cost estimates over the last 24 hours298. Assess monitoring and observability: confirm Cloud Monitoring dashboards, alerting policies, structured logging, OpenTelemetry tracing, and Cloud Error Reporting are configured309. Calculate weighted scores across all categories and determine overall production readiness status3110. Generate a prioritized list of recommendations with estimated score improvement per remediation3233See `${CLAUDE_SKILL_DIR}/references/inspection-workflow.md` for the phased inspection process and `${CLAUDE_SKILL_DIR}/references/inspection-categories.md` for detailed check criteria.3435## Output3637- Inspection report in YAML format with per-category scores and overall readiness percentage38- Runtime configuration summary: model, tools, VPC, scaling settings39- A2A protocol compliance matrix: pass/fail for AgentCard, Task API, Status API40- Security posture score with breakdown: IAM, VPC-SC, Model Armor, encryption, secrets41- Performance metrics dashboard: error rate, latency percentiles, token usage, daily cost estimate42- Prioritized recommendations with estimated score improvement per item4344See `${CLAUDE_SKILL_DIR}/references/example-inspection-report.md` for a complete sample report.4546## Error Handling4748| Error | Cause | Solution |49|-------|-------|----------|50| Agent metadata not accessible | Insufficient IAM permissions or incorrect agent ID | Verify `roles/aiplatform.viewer` granted; confirm agent ID with `gcloud ai agents list` |51| A2A AgentCard endpoint 404 | Agent not configured for A2A protocol or endpoint path incorrect | Check agent configuration for A2A enablement; verify `/.well-known/agent-card` path |52| Cloud Monitoring metrics empty | Monitoring API not enabled or no recent traffic | Run `gcloud services enable monitoring.googleapis.com`; generate test traffic first |53| VPC-SC perimeter blocking access | Inspector running outside VPC Service Controls perimeter | Add inspector service account to access level; use VPC-SC bridge or access policy |54| Code Execution TTL out of range | State TTL set below 1 day or above 14 days | Adjust TTL to 7-14 days for production; values above 14 days are rejected by Agent Engine |5556See `${CLAUDE_SKILL_DIR}/references/errors.md` for additional error scenarios.5758## Examples5960**Scenario 1: Pre-Production Readiness Check** -- Inspect a newly deployed ADK agent before production launch. Run all 28 checklist items across security, performance, monitoring, compliance, and reliability. Target: overall score above 85% before approving production traffic.6162**Scenario 2: Security Audit After IAM Change** -- Re-inspect security posture after modifying service account roles. Validate that least-privilege is maintained (target: IAM score 95%+), VPC-SC perimeter is intact, and Model Armor remains active.6364**Scenario 3: Performance Degradation Investigation** -- Inspect an agent showing elevated error rates. Query 24-hour performance metrics, identify latency spikes at p95/p99, check auto-scaling behavior, and correlate with token usage patterns to isolate the root cause.6566## Resources6768- [Vertex AI Agent Engine Documentation](https://cloud.google.com/vertex-ai/docs/agents) -- deployment and configuration69- [A2A Protocol Specification](https://google.github.io/A2A/) -- AgentCard, Task API, protocol compliance70- [Cloud Monitoring API](https://cloud.google.com/monitoring/api/v3) -- metrics queries and dashboard configuration71- [VPC Service Controls](https://cloud.google.com/vpc-service-controls/docs) -- perimeter setup and access policies72- [Model Armor](https://cloud.google.com/vertex-ai/docs/generative-ai/model-armor) -- prompt injection protection configuration
micsapp