Google Cloud Agent Platform Inference Assistant
Guides developers in connecting to Google Cloud Agent Platform and calling Gemini or OpenMaaS models.
The skill requires interactive confirmation before inference and identifies the model, SDK, and prompt parameters; authentication, token refresh, endpoints, and third-party model calls are disclosed. Points are deducted because sensitive-data handling, least-privilege IAM guidance, input data-flow/retention policy, revocation, and rollback are absent; setup commands also log in, enable an API, and change the environment.
The decision tree and multiple SDK examples make the main paths plausible, and the document discusses token expiry plus common 429, 400, and 404 causes. Points are deducted because the supplied material omits requirements.txt, committed tests, and CI evidence; examples depend on live cloud services, error handling and diagnostics are thin, and verify_all.sh actually issues inference requests.
The skill clearly targets Gemini/OpenMaaS inference, SDK selection, and selected troubleshooting cases, while excluding deployment and evaluation. Points are deducted because trigger boundaries remain limited despite the decision tree, Chinese-language interaction support is not stated, and adaptation to mainland-China network and account availability is unexplained while core use depends on Google Cloud authentication and overseas/global endpoints.
The documentation is layered across safety, setup, decision-making, SDKs, models, and troubleshooting, with examples, limitation notices, and official documentation links; the repository provides Apache-2.0 licensing, support, and contribution paths. Points are deducted because no skill-specific version, changelog, named maintenance responsibility, or update commitment is supplied; the dependency manifest is absent from the provided material, and examples/model lists may become outdated.
The skill supplies directly reusable Python examples, endpoint formats, authentication guidance, model-ID formats, and troubleshooting advice for core inference onboarding. Points are deducted because users must fill in project, region, permissions, and availability details; live-service and model-list instability remains, and there is no statically verifiable representative successful output or complete production-grade retry implementation.
The material includes a revision-pinned source context, code examples, a shell verification script, and several official documentation references, providing some auditability. Points are deducted because this review is static and scripts were not executed; requirements.txt, CI, and test coverage are not supplied, and verify_all.sh is not an offline test, so key-path reproduction is unproven.
- Do not treat the listed model IDs, regions, or SDK behavior as current guarantees; verify official availability, permissions, and project configuration before calling.
- Inference sends prompts and inputs to Google Cloud/the selected model service and may consume quota or incur charges; establish data-governance and redaction controls for sensitive data.
- verify_all.sh performs real inference and changes the local environment; do not run it before confirming cost, quota, and network conditions.
- Mainland-China users should separately validate reachability of the Google Cloud console, authentication endpoints, and global/us-* services.
What does this skill do, and when should you use it?
This is a focused Agent Skill in the google/skills repository for GenAI inference on Google Cloud Agent Platform. It guides developers through Gemini and OpenMaaS calls using the GenAI SDK, OpenAI SDK, or legacy Agent Platform SDK. It covers authentication, global and regional endpoints, model IDs, example models, and troubleshooting for 429, 400, and 404 errors. It does not cover model deployment or model evaluation.
The skill uses the requested model family, SDK, and error details to guide or generate Python inference code. It configures Google Cloud ADC, OAuth access tokens, SDKs, base URLs, project IDs, and regional endpoints. It provides examples for chat completions, legacy text completions, and embeddings, and explains failures caused by DSQ exhaustion, invalid request parameters, unavailable models, or missing Model Garden enablement. Before executing inference, it requires confirmation of parameters such as model ID, SDK, and input prompt.
- A developer starting a Gemini project who wants to use the recommended google-genai SDK.
- An engineer calling Llama, DeepSeek, Qwen, or other OpenMaaS models through a Vertex AI endpoint.
- An application team that wants an OpenAI SDK-compatible interface for Google Cloud model inference.
- A developer troubleshooting 429 Resource Exhausted, 400 User Validation, or 404 Not Found errors.
- A team configuring global or regional Agent Platform endpoints and verifying model IDs and locations.
What are this skill's strengths and limitations?
- Covers both Gemini and OpenMaaS model families.
- Clearly distinguishes the recommended and alternative SDK paths.
- Includes authentication, endpoint configuration, availability checks, and troubleshooting guidance.
- Includes chat-completion, text-completion, and embedding examples.
- The listed model IDs and locations are explicitly examples or incomplete and must be verified.
- Inference can consume cost or quota and requires interactive confirmation.
- It depends on the Google Cloud CLI, a Python environment, network access, and Google Cloud credentials.
- It excludes model deployment and evaluation.
- The source provides no evidence of script test results or a complete test suite.
How do you install this skill?
This skill is part of the google/skills collection. Install the collection with npx skills add google/skills, then select Agent Platform GenAI Inference Skill during installation. The source does not document the exact installation directory.
How do you use this skill?
After installation, use a request such as “Call a DeepSeek OpenMaaS model on Google Cloud Agent Platform with the OpenAI SDK and troubleshoot a 429 error.” Before running Python examples, use gcloud auth login, gcloud auth application-default login, gcloud services enable aiplatform.googleapis.com, python3 -m venv .venv, source .venv/bin/activate, and pip install -r scripts/requirements.txt. Actual inference requires interactive confirmation of the target model, SDK, and input prompt; the source does not include the complete script contents.
How does this skill compare with similar options?
For Gemini, the skill recommends google-genai and presents the legacy vertexai SDK as a still-supported alternative. For OpenMaaS, it highly recommends the OpenAI SDK and presents the GenAI SDK and legacy SDK as alternatives.