Agent Platform DeploySAFE

Agent Skills for Google products and technologies

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: agent-platform-deploy
metadata:
  version: "1.0.4"
  category: AiAndMachineLearning
description: >-
  Deploy open models or custom weights from Model Garden to Agent Platform
  endpoints, check the status of an in-progress deployment operation, or clean
  up resources by undeploying models and deleting endpoints. Use when asked to
  actively deploy a model, list the Model Garden CATALOG of available models,
  check if a specific model is deployable
  (`gcloud ai model-garden models list-deployment-config`), query deployment
  cost, troubleshoot deployment errors (like quota limits), or undeploy/clean
  up endpoints. Also use when copying and deploying a 1P Tuned Model. Don't
  use for pure listing/discovery questions of the form "is X deployed?",
  "list my endpoints", or "which regions have models running?" — for those
  use `agent-platform-endpoint-management`. Don't use for running model evaluations
  (use `agent-platform-eval-flywheel` skill).
---

# Agent Platform Model Garden Deploy Skill

This skill provides instructions for deploying Open Models from Agent Platform
Model Garden to endpoints, and subsequently undeploying them to clean up
resources.

## 1P Tuned Model Copy & Deployment

If you need to copy a **1P (First-Party) Tuned Model** from a source project to
a destination region or project and deploy it to a newly created endpoint, refer
to the
[1P Tuned Model Copy & Deployment Guide](references/copy_deploy_guide.md).

## Safety & Confirmation Tiers (CRITICAL)

Before executing any commands on behalf of the user, you MUST adhere to the
following safety tiers based on the action requested:

1.  **Tier R: Read-only (`list`, `describe`, `list-deployment-config`)**
    *   **Rule**: No confirmation needed. You may execute these commands
        immediately to gather information for the user.
2.  **Tier M: Mutating & Reversible (`deploy`, `undeploy-model`)**
    *   **Rule**: This requires explicit user confirmation. You MUST present a
        clear dry-run confirmation card containing:

        1.  Exact proposed `gcloud` command code block (`gcloud ai model-garden
            models deploy ... --asynchronous`).
        2.  Model identifier, destination project ID, and target region.
        3.  Machine type and accelerator configuration.
        4.  Estimated hourly cost ($/hr).
        5.  Endpoint display name.
        6.  Explicit confirmation prompt asking the user to approve before
            execution.
        You MUST wait for their explicit confirmation before executing. For
        `undeploy-model`, you MUST first verify that the endpoint and deployed
        model exist; if `describe` or `list` returns a 404 or empty result, you
        MUST halt and inform the user rather than attempting undeployment.
    *   **Same-turn restriction**: Do not run the command in the same turn as
        presenting the confirmation prompt. End your turn after asking and wait
        for the user's reply; only execute after explicit approval. Printing a
        preview and then calling the tool before the user can answer does not
        count as obtaining confirmation.
3.  **Tier D: Destructive & Irreversible (`delete`)**
    *   **Rule**: This requires **explicit typed confirmation**. You MUST output
        a text message explaining the irreversible nature of endpoint or model
        deletion and asking the user to type "I confirm" or "Yes, delete it"
        before executing the deletion command.

> [!IMPORTANT]
>
> **Always Output Complete Text Response (NEVER Emit Empty Text)**: After
> executing any tool call (such as `gcloud ai model-garden models deploy`,
> `gcloud ai endpoints delete`, `gcloud ai endpoints list`, or status checks),
> you MUST formulate and return a complete, informative textual response to the
> user. Explicitly report the operation name, endpoint name/ID, error message,
> or list of resources. **NEVER finish a turn with empty text or silence**.

## 1. Prerequisites

Before deploying, ensure you have the correct project and region set. The
commands below use placeholder variables `PROJECT_ID` and `LOCATION_ID`.

Ensure you are authenticated:

```bash
gcloud auth login
gcloud auth application-default login
gcloud config set project $PROJECT_ID
```

## 2. Discovering Deployable Models

You can list models available in Model Garden and check if they can be
self-deployed.

```bash
gcloud ai model-garden models list
```

To see what machine types and accelerators are supported for a specific model,
pass a `MODEL_ID` you obtained from the `models list` output above. Substitute
`<PUBLISHER>/<FAMILY>@<VERSION-ID>` below with the exact string from the catalog
output — the placeholder is deliberately not a real model ID:

```bash
gcloud ai model-garden models list-deployment-config \
    --model="<PUBLISHER>/<FAMILY>@<VERSION-ID>"
```

> [!NOTE] Some models, especially Hugging Face models, might require a Hugging
> Face Access Token for deployment.

> [!TIP] **Model Recommendation Instructions:** Whenever you are about to name a
> specific model version in a response, do NOT recommend from memory. This
> applies in all of the following situations — not just direct deploy requests:
>
> *   The user asks to deploy a model without naming one.
> *   You are volunteering a next-step suggestion after a `list`, `describe`, or
>     `undeploy` operation (e.g. "Would you like me to deploy `<model>` to this
>     endpoint?").
> *   The user asks a general "what should I use?" / "what's a good model for
>     X?" question.
> *   You are filling in a `MODEL_ID` value in an example command you are
>     showing the user (as opposed to a placeholder like
>     `<PUBLISHER>/<FAMILY>@<VERSION-ID>`).
>
> New model versions ship frequently and older ones may be deprecated, so
> training-corpus knowledge of which models exist is unreliable. Follow this
> procedure:
>
> 1.  **Clarify the use case** if it isn't already clear from context (task
>     type, quality vs. latency vs.

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (2)

COPY_RESP=$(curl -s -X POST -H "Authorization: Bearer $(gcloud auth print-access-token)" -H "Content-Type: application/json; charset=utf-8" -d '{ "sourceModel":"'"${MODEL}"'"}' "${ENDPOINT}/v1/project

curl -X POST -H "Authorization: Bearer $(gcloud auth print-access-token)"  -H "Content-Type: application/json" ${ENDPOINT}/v1/${NEW_ENDPOINT}:generateContent -d '{  "contents":  {    "role": "USER",

Gates applied: no_behavioural_pass.

[![ModelsAtlas audit: Agent Platform Deploy](https://modelsatlas.com/skills/google/agent-platform-deploy/badge.svg)](https://modelsatlas.com/skills/google/agent-platform-deploy/)

Not your listing, or something out of date? Claim or update your listing.