Agent Platform Model Deployment Assistant
Deploy, verify, and clean up Model Garden models on Google Agent Platform.
The document separates read-only, reversible, and destructive actions, and requires cost confirmation, typed deletion confirmation, and pre-undeploy existence checks, showing meaningful user-consent and partial recovery controls. Points are deducted because the examples expose a Hugging Face token placeholder, request cross-project IAM changes, and lack least-privilege guidance, sensitive-token handling, end-to-end data-flow disclosure, and a complete recovery plan; cleanup may delete models and endpoints.
The happy path covers discovery, deployment, asynchronous status checks, verification, undeployment, and troubleshooting, with some feedback for long-running operations and errors. Points are deducted because the static files contain no executable tests or CI evidence, several commands and API examples rely on placeholders or incomplete/inconsistent parameters, the copy flow depends on an internal tool, JSON is parsed with grep, and timeout and robust exception handling are absent.
The description clearly covers Model Garden, custom weights, 1P tuned models, status checks, and cleanup, while excluding public Vertex AI deployments and model evaluation. Points are deducted because trigger boundaries, required inputs, output expectations, and permission prerequisites remain incomplete; the material is primarily English and Google Cloud oriented, does not document Chinese interaction, and does not address mainland-China reachability even though the core service and gcloud/curl workflow may depend on overseas services.
The package includes the main skill, a reference guide, a sample prompt, and a configuration script; the README documents installation, Apache-2.0 licensing, active development, and an issue path. Points are deducted because the skill lacks its own version, changelog, named maintenance responsibility, and explicit update path; CLI/API version requirements, dependencies, known limitations, FAQs, and recovery guidance are thin, and the reference contains internal-environment assumptions, hard-coded project data, and unfinished command examples.
The stated scope matches the deployment lifecycle, and the files provide gcloud/curl examples, asynchronous polling, endpoint verification, and cleanup steps that could cover the core task. Points are deducted because users must supply and validate many IDs, permissions, regions, internal endpoints, and model parameters; some commands or request formats are incomplete, leaving substantial manual correction and judgment, and static review cannot verify representative outputs.
The files contain concrete commands, scripts, error conditions, and a fixed revision context, providing limited auditability. Points are deducted because there is no committed test suite, CI execution evidence, real output, or independent review; key API behavior, pricing, and model deployability depend on external links or runtime queries and cannot be independently reproduced from the static material alone.
- Do not place a Hugging Face token, user identity, or project identifier from the examples directly into production; use secure secret injection and least-privilege IAM.
- Before copy, IAM changes, deploy, undeploy, or delete, verify the project, region, endpoint, model, and billing impact item by item; the cleanup commands have incomplete parameters and unclear deletion scope.
- Validate gcloud/API versions, dedicated-endpoint reachability, and deployment configuration first; the material does not establish that the core workflow is usable from mainland-China networks.
What does this skill do, and when should you use it?
This skill guides deployment of open models, custom weights, and 1P Tuned Models to Agent Platform endpoints. It covers model discovery, deployment configuration checks, asynchronous status monitoring, dedicated-endpoint verification, undeployment, and cleanup. Mutating and destructive actions require progressively stronger confirmation, and deployment requires a list-price estimate based on hardware settings. It is not intended for public Vertex AI deployments or model evaluations.
Uses the gcloud CLI to list Model Garden models and inspect deployment configurations; runs gcloud ai model-garden models deploy for Model Garden models or custom weights stored at a GCS URI; checks asynchronous operations with gcloud ai operations describe and endpoints with gcloud ai endpoints list; retrieves dedicatedEndpointDns and sends an authenticated curl request to verify chat serving; lists and describes endpoints and models to identify undeployment resources; performs model undeployment and endpoint/model deletion workflows; and responds to quota or resource-exhaustion failures by asking for confirmation before retrying with alternative regions or hardware.
- A Google Cloud user needs to find a self-deployable Model Garden model and inspect its supported machine types or accelerators.
- A platform engineer needs to deploy an open model or custom weights from GCS to a selected project and region.
- A developer has started an asynchronous deployment and needs to inspect its operation ID or endpoint status.
- An application team needs to send a test prompt to a dedicated endpoint to confirm that serving works.
- A project owner has finished testing and needs to undeploy the model and delete resources to stop ongoing compute charges.
- An operator is troubleshooting QUOTA_EXCEEDED or RESOURCE_EXHAUSTED errors and needs to evaluate an alternative region or machine type.
What are this skill's strengths and limitations?
- Covers the workflow from model discovery through serving verification and cleanup.
- Defines separate confirmation tiers for read-only, reversible, and destructive operations.
- Supports Model Garden models, GCS custom weights, and 1P Tuned Model copy-and-deployment scenarios.
- Includes concrete guidance for dedicated endpoints and quota or resource-exhaustion failures.
- Deployments consume compute resources and incur costs; the source provides pricing guidance but not a complete cost table or calculator.
- Examples depend on the gcloud CLI, authentication, network access, and curl; other client integrations are not documented.
- The source provides no test suite, success-rate data, performance benchmarks, or supported-region matrix.
- The detailed 1P Tuned Model procedure is referenced in `references/copy_deploy_guide.md`, whose contents are not included.
How do you install this skill?
Run npx skills add google/skills, then select the specific skill during installation. The skill is located at skills/cloud/agent-platform-deploy. The source does not document additional installers, version requirements, or an automated test suite.
How do you use this skill?
Provide the project, region, target Model Garden model or custom-weights URI, and hardware requirements. Example trigger: List self-deployable Model Garden models and check the deployment configuration for google/gemma3@gemma-3-27b-it. Read-only list, describe, and list-deployment-config actions require no confirmation. deploy and undeploy-model require explicit confirmation; delete requires typed confirmation such as I confirm or Yes, delete it. Before deployment, the skill must present a prediction list-price estimate based on machine type, accelerator type, and count, noting that discounts or reservations may change the actual bill.
How does this skill compare with similar options?
For public Vertex AI deployments, the source directs users to the vertex-deploy skill. For model evaluations, it directs users to the agent-platform-eval skill. This skill focuses on Agent Platform Model Garden deployment, verification, undeployment, and cleanup.