Google AI Cloud Developer Advocates Detail Graph Workflows in ADK for Executable Processes

Introduction to Graph Workflows in ADK

Graph engineering involves structuring tasks into interconnected nodes and edges, defining the flow of code, models, or human input. The Agent Development Kit (ADK) provides a Workflow feature that translates these designs into executable processes, utilizing functions and agents to perform the work.

This approach allows for the systematic execution of complex operations, as demonstrated through a refund processing example. The workflow can manage parallel steps, route decisions, incorporate human review, and process multiple cases efficiently.

Core Concepts: Fan-out, Fan-in, and Routing

The ADK workflow supports key patterns like fan-out and fan-in. Fan-out enables independent steps to run concurrently, such as multiple data lookups that do not depend on each other's results. Fan-in then consolidates the results from these parallel paths before proceeding to the next stage.

Decision routing can be deterministic or agent-driven, allowing the workflow to adapt based on intermediate results. This flexibility is crucial for handling varied scenarios within a single process.

Human-in-the-Loop and Dynamic Orchestration

Workflows can be configured to include human-in-the-loop pauses, allowing for manual review or intervention at critical junctures. This ensures that complex or sensitive decisions can be vetted by a person before the automated process continues.

The ADK also supports dynamic orchestration, where Python can schedule further work as results become available, offering an alternative to statically defined graph paths. This allows for more adaptive and responsive workflows.

Practical Application: A Refund Workflow Example

A refund request scenario illustrates the practical application of ADK graph workflows. Initially, a single agent could handle the entire request, but breaking it into distinct nodes for lookups, decisions, and responses provides more control. For instance, order, payment, and refund-history lookups can run in parallel since they only require the order ID and are independent of each other.

Once these lookups complete, their results are joined, and a policy decision node can then process the combined information. This structured approach ensures efficiency and clarity in managing complex business logic.