Google ADK 2.0 tutorial: Multi-agent workflows

Google ADK 2.0 workflows are the focus of this tutorial with Miguel Gutierrez and Martin Omander. Gutierrez demonstrates a smart parser that turns unstructured support records into a report. The inputs include email threads, Slack transcripts and PagerDuty logs. He argues that explicit workflow code gives more reliable control over agent coordination than instructions embedded in prompts.

The code review follows the graph through parallel parsing agents, a join node and a report presenter that creates the first draft. A conditional loop then runs a report refiner. Its quality gate checks requirements such as removing confidential links to an internal tracking tool. The example caps refinement at three attempts and stores the attempt count in a context variable. That cap limits repeated work; exiting the loop alone does not establish that every report passes its checks.

Gutierrez also reports an output token limit failure when he tested a thousand cases. He proposes a Python function node to split large inputs into chunks for an AI agent to process in a loop. The tutorial covers workflow structure and safeguards, but does not demonstrate local model execution or a self-hosted deployment.