Graph Workflows in ADK: Everything You Need to Know

The Agent Development Kit (ADK) Workflow transforms graph engineering designs into executable processes by connecting nodes with edges, where functions and agents perform the work. Using a refund example, the system demonstrates parallel execution of independent tasks, such as order, payment, and refund-history lookups, which can run simultaneously since none depend on another’s output. This approach, called fan-out, allows multiple steps to proceed at once, improving efficiency by avoiding unnecessary sequential delays.

The workflow also introduces fan-in, where results from parallel tasks are combined before proceeding. In ADK, lookup functions become nodes directly, and a JoinNode aggregates their outputs into a single dictionary. This structure enables the next node to access all required data without additional model calls, ensuring smooth transitions between steps. Dependencies between nodes are determined by data flow, not the order in which steps are listed in the prompt.

Routing decisions within the workflow are handled by routers, which direct the process based on specific criteria. For refunds, a deterministic router applies fixed policy thresholds to approve, deny, or escalate to manual review. This ensures consistent outcomes for the same case data, allowing policy testing without model calls. Routers can be static, with predefined paths, or dynamic, where an agent interprets intent to select the next step.

Human-in-the-loop pauses are supported, allowing workflows to request human review when necessary. For example, unclear refund requests can trigger follow-up questions, while manual review handles cases that don’t meet automatic approval criteria. The system also supports dynamic orchestration, where Python schedules further work based on intermediate results, offering flexibility in how tasks are executed.