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):
${CLAUDE_SKILL_DIR}/references/implementation.md
- Worked examples (RAG support agent, multimodal extraction):
${CLAUDE_SKILL_DIR}/references/examples.md
- Error-handling and recovery patterns:
${CLAUDE_SKILL_DIR}/references/errors.md
- Product / architecture context:
${CLAUDE_SKILL_DIR}/PRD.md, ${CLAUDE_SKILL_DIR}/ARD.md
- Vertex AI docs: https://cloud.google.com/vertex-ai/docs
- Agent Engine docs:
1---2name: vertex-agent-builder3description: 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). Use when designing, deploying, or operating Vertex AI agents on Google Cloud. Trigger with "build a Vertex agent", "deploy to Agent Engine", or "wire up RAG on Vertex AI".4license: MIT5---6# Vertex AI Agent Builder
7
8Build and deploy production-ready agents on Vertex AI with Gemini models, retrieval (RAG), function calling, and operational guardrails (validation, monitoring, cost controls).
9
10## Overview
11
12- Produces an agent scaffold aligned with Vertex AI Agent Engine deployment patterns.
13- Helps choose models/regions, design tool/function interfaces, and wire up retrieval.
14- Includes an evaluation + smoke-test checklist so deployments don’t regress.
15
16## Prerequisites
17
18- Google Cloud project with Vertex AI API enabled
19- Permissions to deploy/operate Agent Engine runtimes (or a local-only build target)
20- If using RAG: a document source (GCS/BigQuery/Firestore/etc) and an embeddings/index strategy
21- Secrets handled via env vars or Secret Manager (never committed)
22
23## Instructions
24
251. Clarify the agent’s job (user intents, inputs/outputs, latency and cost constraints).
262. Choose model + region and define tool/function interfaces (schemas, error contracts).
273. Implement retrieval (if needed): chunking, embeddings, index, and a “citation-first” response format.
284. Add evaluation: golden prompts, offline checks, and a minimal online smoke test.
295. Deploy (optional): provide the exact deployment command/config and verify endpoints + permissions.
306. Add ops: logs/metrics, alerting, quota/cost guardrails, and rollback steps.
31
32## Output
33
34- A Vertex AI agent scaffold (code/config) with clear extension points
35- A retrieval plan (when applicable) and a validation/evaluation checklist
36- Optional: deployment commands and post-deploy health checks
37
38## Error Handling
39
40- Quota/region issues: detect the failing service/quota and propose a scoped fix.
41- Auth failures: identify the principal and missing role; prefer least-privilege remediation.
42- Retrieval failures: validate indexing/embedding dimensions and add fallback behavior.
43- Tool/function errors: enforce structured error responses and add regression tests.
44
45## Examples
46
47**Example: RAG support agent**
48
49- Request: “Deploy a support bot that answers from our docs with citations.”
50- Result: ingestion plan, retrieval wiring, evaluation prompts, and a smoke test that verifies citations.
51
52**Example: Multimodal intake agent**
53
54- Request: “Build an agent that extracts structured fields from PDFs/images and routes tasks.”
55- Result: schema-first extraction prompts, tool interface contracts, and validation examples.
56
57## Resources
58
59- Implementation patterns (model selection, RAG wiring, deployment config): `${CLAUDE_SKILL_DIR}/references/implementation.md`
60- Worked examples (RAG support agent, multimodal extraction): `${CLAUDE_SKILL_DIR}/references/examples.md`
61- Error-handling and recovery patterns: `${CLAUDE_SKILL_DIR}/references/errors.md`
62- Product / architecture context: `${CLAUDE_SKILL_DIR}/PRD.md`, `${CLAUDE_SKILL_DIR}/ARD.md`
63- Vertex AI docs: https://cloud.google.com/vertex-ai/docs
64- Agent Engine docs:
gabrielmoreira