How to use the Agent Platform Alert Setup agent skill

What does the skill make the agent do?

When you ask your AI assistant to set up monitoring for an AI agent you run on Google Cloud, the skill makes it follow a fixed process. It first checks that the agent is emitting telemetry, then looks for existing alerts so it does not create duplicates, and then writes alert rules as Terraform configuration files.

The rules cover latency, errors, cost, safety, security and, for Vertex AI agents, quality. The skill is built around OpenTelemetry (OTel) metrics, so the agent must already emit those metrics for the alerts to have data to evaluate.

• Reliability: latency, fast and slow burn error-rate SLOs, model call error rate, tool call error rate
• Quality (Vertex AI agents only): final response quality, tool use quality, hallucination
• Cost: rapid token burn rate
• Safety: high Model Armor safety policy trigger rate
• Security: high IAM permission denied trigger rate

Install the skill from the ZIP or the Claude apps

You can install the skill in one of two ways. For Claude Code, download the ZIP from the skill's page on this directory, then unzip it into ~/.claude/skills/ to make it available to all your projects, or into a project's .claude/skills/ folder to limit it to that project.

In the Claude apps, download the same ZIP and upload it under Settings → Capabilities → Skills.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Set up the prerequisites

The skill's scripts need Python 3 and a few client libraries. Install them from the requirements file that ships with the skill before running any script. The skill also needs the gcloud CLI with credentials that can access your target project, and Terraform to apply the generated configuration.

Your agent must be instrumented with OpenTelemetry metrics. Without them, the Reliability, Cost, Safety and Security policies have no data to evaluate.

• Python 3 with the dependencies in scripts/requirements.txt (Google Cloud monitoring and aiplatform client libraries, google-auth)
• gcloud CLI with Google Cloud credentials for the target project
• Terraform for the generated configuration
• An agent instrumented with OpenTelemetry metrics

Trigger the skill with a natural request

You do not need to type a special command. Describe the outcome you want and name the agent and project. The skill is meant to be used when you are analyzing, writing or deploying alerting policies for an agent's latency, error rates, token usage or quality.

The skill only works on the projects you name in the request. It will not reuse a project from earlier in your session unless you direct it to.

• "Set up alerting for my agent support-bot in project my-project."
• "Check whether my Vertex AI agent is emitting OpenTelemetry metrics, then write alert policies for latency and errors."
• "Write Terraform alerts for token usage spikes on agent billing-assistant in project my-project."

What happens during a run

The skill runs a discovery script first. It identifies the agent runtime, verifies telemetry, and finds metric scopes and linked datasets. If that fails or returns partial data, the assistant falls back to manual checks: it lists metric scopes with gcloud, searches Terraform for scoping projects, and asks you if the scoping project is still unclear.

Before writing anything, the assistant scans for existing policies that target the same metrics. Unless you ask for specific alert types, it configures all of them, and it writes the output only as Terraform.

Safety notes: billing, scripts and network access

Read-only checks run without confirmation. Anything that creates billable resources needs your explicit approval. The assistant must warn you about extra costs from LLM evaluations for the online monitor, and about Cloud Trace or Cloud Logging export for telemetry, and must stop and ask before provisioning.

The skill runs scripts and uses the network. Review the scripts and the generated Terraform before applying it. Do not apply the output to production without review, because it creates cloud resources and charges.

• Read-only scripts run without confirmation
• Provisioning requires your explicit approval
• Review all generated Terraform before applying it to production

Tips for better results

Name one project and one agent per request. If you run several agents in a Cloud Monitoring metrics scope, say so, so the policies go to the scoping project. Quality alerts only work for Vertex AI deployments, so tell the assistant where your agent runs.

• Give the project ID and agent name in the same message
• Say if you use a multi-project metrics scope and name the scoping project
• Ask for specific alert types if you do not want all of them
• Check telemetry first with a read-only request