GCP Best Practices — propane-will-491900-m5
Platform context (2025–2026)
- Vertex AI Agent Engine has been rebranded as Gemini Enterprise Agent Platform.
Agent runtimes are fully managed serverless — no VM management required.
- ADK (Agent Development Kit) 2.0 is open-source, supports Python/TypeScript/Go/Java/Kotlin.
Agents run locally for dev, then deploy to Agent Runtime, Cloud Run, or GKE.
- Region: us-central1 for all resources (low latency, widest feature availability).
Required APIs (enable once, idempotent)
aiplatform.googleapis.com # Vertex AI / Gemini Agent Platform
storage.googleapis.com # Cloud Storage staging
cloudbilling.googleapis.com # Billing Budget API
cloudresourcemanager.googleapis.com
iam.googleapis.com
run.googleapis.com # If using Cloud Run as deploy target
monitoring.googleapis.com
logging.googleapis.com
cloudtrace.googleapis.com
secretmanager.googleapis.com
artifactregistry.googleapis.com
cloudbuild.googleapis.com
ADK install
pip install google-adk>=2.0.0 google-cloud-aiplatform[adk,agent_engines]>=1.112.0
gcloud components update
gcloud auth application-default login
Deploy targets — decision matrix
| Target |
Best for |
Cost model |
Auth |
| Agent Runtime (managed) |
Production agents, Memory Bank, Sessions |
Per vCPU-second + GiB-second |
ADC / SA |
| Cloud Run |
Custom containers, public API endpoints, lower ops |
Pay-per-request, scales to 0 |
ADC / SA |
| GKE |
Complex multi-agent meshes, custom networking |
Cluster uptime + node cost |
Workload Identity |
ADK 2.0 — new in 2026
- Graph workflows and collaborative agents (coordinator + sub-agents)
- ADK Kotlin — new language support
adk deploy cloud_run and adk deploy agent_engine CLI commands- Native MCP client and server support
- A2A (Agent-to-Agent) protocol support
SecretManagerClient integration (requires ADK ≥ 1.29)
Four platform pillars (Gemini Enterprise Agent Platform)
| Pillar |
Key features |
| Build |
ADK 2.0, Agent Studio (visual designer), Agents CLI, 200+ models via Model Garden |
| Scale |
Agent Runtime, Sessions, Memory Bank, Code Execution, multi-day workflows |
| Govern |
Agent Identity, Agent Registry, Agent Gateway, Model Armor, Agent Security dashboard |
| Optimize |
Agent Evaluation, Agent Optimizer (auto-clusters failures), Example Store, Observability |
Model landscape (May 2026)
| Model |
Status |
Recommended use |
gemini-2.5-flash |
GA |
Balanced cost/quality — recommended default |
gemini-3-flash |
Public preview |
Cheapest option for dev/test |
gemma-4 |
GA (open-source) |
Self-hosted / fine-tuning |
gemini-1.5-* |
Discontinued |
Migrate immediately |
claude-3.5-haiku |
Deprecated (shutdown Jul 2026) |
Migrate immediately |
agent-starter-pack (Google official)
uvx agent-starter-pack enhance # add CI/CD + eval + observability to existing agent
uvx agent-starter-pack create my-agent # scaffold from scratch
Provides: CI/CD (Cloud Build), Evaluation, Observability, Terraform IaC out of the box.
GitHub: https://github.com/GoogleCloudPlatform/agent-starter-pack
Key constraints
- Never use GPU instances (A100/V100) unless business-critical.
- Default compute:
e2-micro for background tasks, up to e2-standard-4 for Colab.
- Agent Runtimes bill when deployed — always run
03-teardown.sh at end of day.
- Cloud Storage buckets do NOT auto-delete — lifecycle rules are mandatory.
