Scaffold Prod AI Agents on GCP in 60 Seconds
Agent Starter Pack generates full production infrastructure (CI/CD, Terraform, eval, observability) around any agent framework via one CLI command and 6 templates, slashing 3-9 months of setup—but GCP-only with no official support.
Generate Complete Production Stack in One Command
Run uvx agent-starter-pack create and answer two prompts (template and deployment target) to output a full project in 60 seconds: FastAPI backend with auth, React chat frontend, Terraform for GCP resources, CI/CD via Cloud Build/GitHub Actions, Vertex AI evaluation, Cloud Logging/Trace observability, and auto-docs. This eliminates weeks of manual YAML/Terraform boilerplate, wiring eval before deployment to catch laptop-to-prod failures that trap most teams for 3-9 months.
The stack addresses four core gaps: customization with real data (vector search), evaluation (Vertex AI metrics pre/post-deploy), deployment (scalable infra/CI/CD), and observability (real-time traces/logs/Looker dashboards). Unlike frameworks like LangGraph (verbose state schemas/nodes/edges) or CrewAI (no checkpointing for long runs), it wraps your chosen orchestration—Google AI/ADK/A2A/LangGraph—focusing on surrounding infra.
Six Templates Match Agent Use Cases
Choose from production-ready starters, all sharing identical infra:
- ADK: Base ReAct agent via Google's Agent Development Kit.
- ADK + A2A: Adds agent-to-agent protocol for cross-framework coordination (e.g., ADK invokes LangGraph agents via standardized tasks).
- Agentic RAG: Vertex AI search/vector Q&A on docs.
- LangGraph: ReAct with LangChain state persistence.
- ADK Java: ReAct for Java teams.
- ADK Live: Real-time multimodal (audio/video/text) via Gemini.
Repo has 6,000 GitHub stars, 1,400 forks, weekly releases for a year, Apache 2.0 license, from Google Cloud team. Deployment targets: Cloud Run (containerized, pay-per-use scaling/control) or Agent Engine (managed runtime/security/VPC compliance). Switch with one flag.
Trade-offs: Power with GCP Lock-in
Vertex AI eval runs quality checks automatically; traces capture request paths; logs enable real-time search. A2A enables multi-agent systems without framework silos. But it's unofficial ("demonstrative," no SLAs/support), Python-first (Java template limited), and GCP-exclusive (Vertex/Cloud Run/Agent Engine—no AWS/Azure). Running infra incurs costs. Use if GCP-aligned and prioritizing speed over flexibility; otherwise, build custom to avoid vendor lock.