Google Cloud BlogAnnie Wang1 min readtutorialintermediate
Graph Workflows in ADK: Everything You Need to Know
Summary
The post walks through building a refund processing workflow with Google’s Agent Development Kit, covering parallel execution, routing decisions, human approvals, and when to use static versus dynamic graph definitions. It includes concrete Python snippets showing fan‑out/fan‑in and router patterns.
- ADK workflows let you model a process as a directed graph of agents and functions.
- Use fan‑out nodes to run multiple refund checks in parallel and fan‑in nodes to aggregate results.
- Deterministic routers choose a static next step, while agent routers can invoke LLMs to decide dynamically.
- Human‑in‑the‑loop steps pause the graph until an operator approves or adds data.
Engineers building AI‑augmented pipelines on Google Cloud should know how to structure and orchestrate tasks efficiently with ADK.
6/10



