Build and deploy generative AI agents on Vertex AI: Gemini model selection, RAG with grounded retrieval, function calling, multimodal extraction, evaluation, and Agent Engine deployment with operational guardrails (logs, alerts, cost controls).
ReadWriteEditGrepBash(cmd:*)
Vertex AI Agent Builder
Build and deploy production-ready agents on Vertex AI with Gemini models, retrieval (RAG), function calling, and operational guardrails (validation, monitoring, cost controls).
Overview
- Produces an agent scaffold aligned with Vertex AI Agent Engine deployment patterns.
- Helps choose models/regions, design tool/function interfaces, and wire up retrieval.
- Includes an evaluation + smoke-test checklist so deployments don’t regress.
Prerequisites
- Google Cloud project with Vertex AI API enabled
- Permissions to deploy/operate Agent Engine runtimes (or a local-only build target)
- If using RAG: a document source (GCS/BigQuery/Firestore/etc) and an embeddings/index strategy
- Secrets handled via env vars or Secret Manager (never committed)
Instructions
- Clarify the agent’s job (user intents, inputs/outputs, latency and cost constraints).
- Choose model + region and define tool/function interfaces (schemas, error contracts).
- Implement retrieval (if needed): chunking, embeddings, index, and a “citation-first” response format.
- Add evaluation: golden prompts, offline checks, and a minimal online smoke test.
- Deploy (optional): provide the exact deployment command/config and verify endpoints + permissions.
- Add ops: logs/metrics, alerting, quota/cost guardrails, and rollback steps.
Output
- A Vertex AI agent scaffold (code/config) with clear extension points
- A retrieval plan (when applicable) and a validation/evaluation checklist
- Optional: deployment commands and post-deploy health checks
Error Handling
- Quota/region issues: detect the failing service/quota and propose a scoped fix.
- Auth failures: identify the principal and missing role; prefer least-privilege remediation.
- Retrieval failures: validate indexing/embedding dimensions and add fallback behavior.
- Tool/function errors: enforce structured error responses and add regression tests.
Examples
Example: RAG support agent
- Request: “Deploy a support bot that answers from our docs with citations.”
- Result: ingestion plan, retrieval wiring, evaluation prompts, and a smoke test that verifies citations.
Example: Multimodal intake agent
- Request: “Build an agent that extracts structured fields from PDFs/images and routes tasks.”
- Result: schema-first extraction prompts, tool interface contracts, and validation examples.
Resources
- Implementation patterns (model selection, RAG wiring, deployment config):
${CLAUDESKILLDIR}/references/implementation.md
- Worked examples (RAG support agent, multimodal extraction):
${CLAUDESKILLDIR}/references/examp