Google ADK 2.0 Alpha: agent teams and operating modes

This overview explains collaborative agent teams in Google's Agent Development Kit (ADK) 2.0 Alpha 1. The speaker describes ADK as an open source framework for multi-agent systems and contrasts graph-based workflows with tasks whose steps need more flexible coordination.

An illustrative travel-planning example explains how a coordinator delegates work through routing tools that ADK automatically injects. A weather checker runs in single-turn mode, fetches weather through geocoding and weather tools, and returns its result without user interaction. A flight booker uses task mode to ask clarifying questions before returning a booking confirmation. An itinerary agent uses chat mode for an ongoing conversation, then explicitly hands control back to the coordinator.

The speaker says chat is the default mode. Chat and task modes do not support parallel execution in the explanation, while single-turn mode does. The video also describes separate session branches for parallel work: subagents cannot see their peers' activity, and the parent gathers their results before deciding what to do next.

For developers designing an AI agent team, the release limitation matters: task agents must be leaf nodes and cannot have subagents of their own in this Alpha release. The presentation explains coordination behavior; it does not demonstrate how to run models locally or establish offline support.